NeuralWired’s Technology section covers the developments reshaping how the world builds, deploys, and regulates digital innovation. We report daily on the stories driving global conversation in artificial intelligence, big technology companies, startups and venture funding, cybersecurity, consumer gadgets and devices, and blockchain and cryptocurrency.
Our technology coverage goes beyond product announcements. When a major AI model launches, we explain what it can actually do and where its claims are overstated. When a startup raises a large funding round, we look at whether the business behind it can sustain that valuation. When a cybersecurity breach hits the news, we explain who is affected and what comes next, not just what happened. Each article is built from original research into primary sources, including company statements, technical documentation, regulatory filings, and verified data, and is written by our editorial team rather than generated automatically.
Readers come to this section for daily updates on the technology stories that matter globally, from shifts inside major technology companies to emerging tools changing how people work, communicate, and build. Whether you are a founder, an investor, an engineer, or simply someone trying to understand where technology is heading next, NeuralWired’s Technology coverage is built to keep you informed without wasting your time on hype.
Best LLM Evaluation Tools 2026: 7 Tested, Ranked, Compared
ML Tooling / Developer Focus
Best LLM Evaluation Tools 2026: 7 Tested, Compared
By the NeuralWired ML Team · Updated July 23, 2026 · Part of our ongoing Tested series
Your RAG pipeline passed every demo. Then it hit production and started citing the wrong policy document to real customers, and nobody noticed for six days. That gap between “looked fine in the demo” and “actually correct at scale” is exactly why best LLM evaluation tools 2026 has become one of the most searched phrases among ML teams this year. We tested seven of them against real traces, real budgets, and real failure modes, so you don’t have to guess which one fits your stack.
Three things collided in 2026 to push evaluation tooling from “nice to have” to “line item in the budget.” First, the money: Gartner’s own newsroom forecasts the global GenAI models market will top 25 billion dollars in 2026 and reach 75 billion by 2029, with LLM observability investment climbing to half of all GenAI deployments by 2028, up from roughly 15 percent in early 2026. Second, the regulation: EU AI Act enforcement begins in August 2026, and several compliance analysts now cite documented evaluation practice as a direct requirement for systems serving EU users, not a suggestion.
Third, and this is the one engineers actually feel: the classic benchmarks stopped telling you anything. MMLU and HellaSwag scores for frontier models now cluster above 88 percent, close enough that the differences sit inside measurement noise. If you’re still leaning on those numbers to pick a model or judge, you’re comparing rounding errors. That’s part of why harder, contamination-resistant suites like GPQA-Diamond, SWE-bench Verified, and LiveBench have taken over as the benchmarks that actually separate models.
The number to budget against: running DeepEval and RAGAS together across 10,000 RAG traces a day costs roughly 200 to 600 dollars a month in LLM-judge token spend on GPT-4o, according to a cost-modeling analysis from genai.qa. Swap in a cheaper judge like GPT-4o-mini or Claude Haiku and that drops 60 to 80 percent.
And here’s the part most “best tools” roundups skip: plenty of teams still haven’t automated this at all. A LangChain survey of more than 1,300 practitioners found 59.8 percent still rely on human review to grade outputs, while 53.3 percent use LLM-as-a-judge, with heavy overlap between the two. The share of teams doing no testing whatsoever did drop, from 29.5 percent to 22.8 percent, but that’s still nearly a quarter of production teams flying blind.
The 7 tools, ranked and compared
We grouped these by what they’re actually built for, because “best overall” is the wrong question. A code-first library that lives in your CI pipeline solves a different problem than a hosted platform your product team logs into. Picking between them is now an architecture decision made at design time, not a bolt-on before launch.
Tool
Best for
Model
Standout feature
DeepEval
CI/CD test suites, pytest-style workflows
Open source
50+ metrics, built by Confident AI’s Jeffrey Ip and Kritin Vongthongsri
RAGAS
RAG-specific scoring
Open source
Reference-free faithfulness, relevancy, and context metrics; academic roots
Promptfoo
Red-teaming, prompt regression
Open source
YAML config, 40+ adversarial testing plugins
Langfuse
Production tracing, self-hosted teams
Open source (MIT), acquired by ClickHouse
Free self-hosted tier, no usage cap
Arize Phoenix
OpenTelemetry-native tracing
Open source + commercial (Arize AX)
Built by former Uber Michelangelo lead Aparna Dhinakaran’s team
Braintrust
Full eval-to-production CI platform
Commercial
Free tier: 1M spans/month, 10K evals
Galileo AI
Regulated, high-volume production
Commercial
Luna-2 judge models for sub-200ms scoring at scale
DeepEval vs RAGAS: the comparison everyone actually searches for
RAGAS is the specialist. It was built for one job, scoring retrieval-augmented generation, and it does that job with four metrics that don’t require a fixed reference answer: faithfulness, answer relevancy, context precision, and context recall. It traces back to a 2023 research paper that reportedly got a mention from OpenAI at a developer event that year, which is part of why it still carries more academic weight than most commercial entrants launched since.
DeepEval is the generalist. It handles RAG too, but also agents, chatbots, and general-purpose CI/CD test gates, all in a pytest-style workflow your existing test suite already understands. If you’re only doing RAG, RAGAS is narrower and arguably sharper. If you’re shipping agents and chatbots alongside RAG, DeepEval covers more ground without forcing you into three separate frameworks. Most teams we found running mature pipelines use both, RAGAS for the retrieval layer, DeepEval for everything wrapped around it.
The observability layer: Langfuse vs Arize Phoenix
Langfuse got acquired by ClickHouse in January 2026, and the open-source repo has kept shipping since, with the self-hosted, MIT-licensed version still free and uncapped. Phoenix, Arize’s open-source tracing library, is OpenTelemetry-native, which matters if your infrastructure team already standardized on OTel for everything else. One licensing note worth flagging honestly: sources describe Phoenix’s license inconsistently, some say Apache 2.0, others Elastic License 2.0, so confirm the current terms directly against the Arize-ai/phoenix repository before you build a dependency on it.
How accurate is LLM-as-a-judge, really
Nearly every tool on this list leans on the same underlying mechanism: one model scoring another’s output against a rubric instead of a fixed correct answer. Aggregated studies put that agreement with human raters somewhere between 80 and 92 percent, at a fraction of the cost of full manual review, roughly 500 to 5,000 times cheaper depending on the study.
That’s genuinely good. It is not perfect, and treating it as a solved problem is where teams get burned. Position bias, verbosity bias (judges tend to prefer longer answers even when they’re not better), and self-preference bias, where a judge model rates outputs from its own model family more favorably, are all documented and unresolved.
“Marketing, not science.”
Nathan Lambert, AI researcher, Interconnects.ai, on cross-vendor benchmark comparisons that can’t be independently verified. Read the original piece
Lambert’s line is from December 2023, and it’s still getting cited in curated 2026 evaluation reading lists for a reason. His point wasn’t about the tools on this list specifically, it was about competitors publishing benchmark claims about each other’s models without access to verify them. The same skepticism applies here: DeepEval, Braintrust, and Latitude all publish comparison content that ranks their own product first. Worth remembering while you read anyone’s “best tools” list, including this one.
What Husain and Dhinakaran are actually saying
Hamel Husain, the independent ML consultant co-authoring the upcoming O’Reilly book on AI evaluation with Shreya Shankar, has argued in multiple interviews that teams who appear to ship without formal evals are usually leaning on evaluation work someone else already did upstream, most often the model provider’s own internal testing. His and Shankar’s broader position, laid out across their public evals masterclass materials, warns specifically against fully automating evaluation without keeping a human grounded in product-specific context.
Aparna Dhinakaran, Arize’s co-founder and chief product officer, framed the shift differently at Arize’s Observe 2026 event: the industry is moving away from a person manually reading individual traces one by one, toward a person overseeing a fleet of agents that check each other’s work. It’s a more optimistic read than Husain’s, and both are right about different parts of the pipeline.
The stack recipes teams actually run
No single tool here covers RAG, agents, and chatbots equally well. The category is fragmented on purpose, and stacking two or three tools for different failure modes is standard practice among teams that have been doing this for a while, not a sign that someone picked wrong the first time. The pattern we saw repeated most often:
CI gate: DeepEval or Promptfoo, catching regressions before merge
RAG-specific scoring: RAGAS, layered in wherever retrieval is involved
Production sampling: Langfuse or Phoenix, watching what real traffic actually does
That combination gets most of a full commercial platform’s coverage for a fraction of the price. If your team is already dealing with agent deployments that keep failing in production, this is the stack worth building before you shop for anything commercial.
Where the hype outruns the evidence
A few things this space glosses over that are worth saying plainly. First, benchmark saturation is real, not a talking point. Second, an 80 to 92 percent agreement rate with human raters means a judge model is wrong on a meaningful slice of calls, and none of the vendor pages selling “automated evaluation at scale” put that number front and center. Third, the timeline claim that any single tool “solves” LLM evaluation is overstated across nearly every vendor’s own marketing page, including the ones cited in this article.
Our read: the risk scenario worth planning around isn’t picking the wrong tool. It’s picking a commercial platform, trusting its automated pass rate completely, and losing the human error-analysis step that would have caught something product-specific a rubric never would.
FAQ
What is the best LLM evaluation tool in 2026?
There’s no single best tool, it depends on use case. DeepEval and Promptfoo lead for CI/CD-integrated testing, RAGAS is the standard for RAG-specific metrics, and Langfuse or Arize Phoenix lead production tracing. Most mature teams combine two or three tools rather than relying on one platform.
Is RAGAS or DeepEval better for RAG evaluation?
RAGAS is purpose-built for RAG with four research-backed reference-free metrics: faithfulness, answer relevancy, context precision, and context recall. DeepEval covers RAG plus agents, chatbots, and CI/CD integration, making it broader but less specialized. Many teams pair both.
How much does LLM evaluation cost at scale?
Running DeepEval and RAGAS together against 10,000 RAG traces a day costs roughly 200 to 600 dollars a month in LLM-judge API spend on GPT-4o. Switching to cheaper judge models like GPT-4o-mini or Claude Haiku cuts that by 60 to 80 percent, with a manageable accuracy trade-off.
What is LLM-as-a-judge and how accurate is it?
LLM-as-a-judge uses one LLM to score another’s output against a rubric instead of comparing it to a fixed answer. Published studies put agreement with human raters at roughly 80 to 92 percent, at a fraction of the cost of full human review, but it carries known biases like favoring longer or first-listed answers.
Are MMLU and other classic benchmarks still useful in 2026?
Only marginally for frontier models. MMLU and HellaSwag scores have saturated above 88 percent, with differences between top models falling inside measurement noise. Harder, contamination-resistant benchmarks like GPQA-Diamond, SWE-bench Verified, and LiveBench are now more informative for comparing leading LLMs.
Is Langfuse free?
Langfuse’s self-hosted version is free and MIT-licensed with no usage limits. The managed cloud version has a free Hobby tier of 50,000 units a month, then paid tiers starting around 29 to 199 dollars a month depending on retention and support needs.
Where this goes next
What you now know that most “top 7 tools” listicles won’t tell you: there is no finish line here. The tools on this list are converging, ClickHouse now owns Langfuse, commercial platforms keep absorbing open-source primitives, and the EU AI Act’s August enforcement date is going to force teams that have been skipping documented evaluation to catch up fast.
Three things worth watching over the next 6 to 18 months:
Whether Galileo and Braintrust’s judge-model approach (cheaper, faster, purpose-built judges instead of general-purpose GPT-4o calls) becomes the default rather than the exception
Whether agent-specific, multi-turn evaluation standards mature enough to replace the single-turn accuracy checks most of this tooling was originally built around
Whether EU AI Act enforcement actually changes tool selection, or just adds a compliance checkbox on top of whatever teams were already running
If you’re picking a stack this quarter: start with the CI gate, add the RAG-specific layer only if you’re running RAG, and don’t buy a commercial platform until you’ve felt the limits of the free, open-source combination first.
Nearly 3 in 10 Small Businesses Hit by Deepfake Scams in 2026
NeuralWired Cybersecurity Desk · Published July 23, 2026
In February 2024, a finance employee at UK engineering firm Arup joined what looked like a routine video call with the CFO and several colleagues. He wired $25.6 million across 15 transactions before anyone realized every face on that call except his own was AI generated. Two years later, that trick has trickled all the way down to businesses with a dozen employees and no IT department: 29% of small businesses now say they’ve experienced a deepfake scam in the past year, according to a new survey from cybersecurity firm VikingCloud.
That number, buried inside VikingCloud’s 2026 SMB Threat Landscape Report, is the clearest signal yet that deepfake fraud stopped being an enterprise problem sometime in the last eighteen months. It’s now a Tuesday-afternoon problem for a plumbing company in Ohio or a marketing agency in Manchester. And the FBI, for the first time in its Internet Crime Complaint Center’s roughly 25-year history, agrees the threat is big enough to track on its own.
VikingCloud surveyed small business owners and operators for its 2026 threat report, and the results reorder what SMBs are worried about. More than a quarter said they’d experienced a deepfake scheme (29%), a customer data breach (27%), a ransomware attack (26%), or a denial of service attack (26%) in the past year. Taken together, 75% of SMB owners now rank cyberattacks as their number one operational threat for 2026, the first time in this survey series that cybersecurity has outranked economic pressure. Forty percent said a cyberattack costing $100,000 or less could put them out of business entirely.
A note on the source VikingCloud hasn’t published full survey methodology, sample size, or margin of error in its public summary; the underlying data sits behind a lead-gen form. That doesn’t make the 29% figure false, but it means it should be read as “according to a vendor survey of small business owners,” not as census-grade data. Compare it against the FBI figure below, which is independently audited.
The FBI Just Made It Official
For 2025, the FBI’s Internet Crime Complaint Center broke out AI-enabled fraud as its own standalone category for the first time. IC3 logged 22,364 complaints with a reported AI nexus, totaling $893,346,472 in adjusted losses, according to the FBI IC3 2025 Annual Report published in April 2026. That figure is the closest thing this space has to a government-audited number, and it’s worth breaking down by category.
Fraud category (AI referenced)
2025 adjusted losses
Investment fraud
$632.0 million
Business email compromise
$30.3 million
Tech and customer-support scams
$19.5 million
Confidence and romance scams
$19.0 million
Employment scams
$12.6 million
Business email compromise is the line that should matter most to a small business owner. It’s the category built entirely around impersonating someone the victim already trusts, a vendor, a boss, a bank contact, and it’s exactly the mechanism behind the Arup case.
The Case That Changed Everything: Arup’s $25.6 Million Call
Arup’s Hong Kong finance team received what appeared to be a standard request from the company’s UK-based CFO: move funds for a confidential transaction. The employee had doubts, so he did what security training tells you to do. He joined a video call to verify. Every other participant on that call, including the person who looked and sounded like the CFO, was an AI-generated deepfake. He made the transfers. Reporting from the Financial Times and CNN in May 2024 confirmed the total loss at $25.6 million across 15 wire transactions, and the case has become the reference point every security vendor cites when explaining why video verification alone is no longer enough.
Arup is a global engineering firm with sophisticated finance operations, not a small business. That distinction matters, and we’ll come back to it. But the mechanics of the attack, real-time video and voice synthesis convincing enough to fool someone who was actively trying to verify, work exactly the same way against a five-person accounting team as they did against Arup’s.
When the Defense Works: WPP’s Near Miss
Not every attempt succeeds, and the counter-example is worth knowing. Scammers targeted WPP CEO Mark Read using a cloned voice and a spoofed Microsoft Teams meeting invite, built around a fake WhatsApp account using his public photo, according to an entry in the OECD.AI Incident Database and reporting from Marketing-Interactive. Staff escalated before any money moved. WPP confirmed zero losses.
What stopped it wasn’t detection software. It was a human asking a question the scammer couldn’t answer and refusing to proceed until someone verified through a separate channel. That’s a cheap lesson, and it’s the same one at the center of the advice section below.
Why Small Businesses Are the Easier Target
Here’s the uncomfortable part for small business owners: being small isn’t protection. It’s the opposite. VikingCloud’s data shows 84% of SMB owners self-manage their own cybersecurity, with no dedicated IT or security staff. That means the same person approving a vendor invoice is also the last line of defense against a fraudulent one, with no gatekeeper, no second sign-off, no layered approval chain to slow things down.
An enterprise like Arup still has structural weaknesses attackers can exploit, but it also has finance controls, compliance teams, and escalation paths. A twelve-person business usually has one bookkeeper and a Slack channel. Attackers know which door is easier to walk through.
“We only have like one really good example in the news right now of that organization in Hong Kong that ended up falling for and sending $25 million based on a deepfake audio and video scam, and I think we’re going to see a lot more business email compromise style events because of AI.”Rachel Tobac, CEO, SocialProof Security · 8th Layer Insights podcast, The Cyber Wire, April 9, 2024
Tobac’s prediction has aged into the current data. The FBI’s BEC-with-AI-nexus figure alone hit $30.3 million in 2025, and that’s before counting the cases that never get formally reported, which fraud researchers generally assume is the majority of them.
“AI-generated media is not just a future risk, it’s a real business threat. We’re seeing executives impersonated, hiring processes compromised, and financial safeguards bypassed with alarming ease.”Tony Lee, Head of Consulting, Hong Kong & Macau, Trend Micro · Media OutReach Newswire, July 10, 2025
Worth flagging: Lee’s employer, Trend Micro, sells deepfake detection tools, so treat the quote as an informed but interested voice rather than a neutral one.
Can You Trust Your Own Eyes?
Most SMB owners assume they’d notice if something felt off on a call. The data says otherwise. Controlled lab studies compiled by security research firm DeepStrike found human accuracy at spotting high-quality deepfake video sits at just 24.5%, even though roughly 60% of people believe they could identify one. That gap between confidence and competence is arguably the more dangerous number in this whole story.
The technical barrier to producing convincing fakes keeps dropping too. McAfee’s consumer research found a voice clone with about 85% similarity to the original can now be generated from just three seconds of audio, easily pulled from a podcast clip, a local news interview, or a company’s own marketing video.
The Regulatory Clock Is Ticking
Two regulatory shifts land right around this article’s publish date. The EU AI Act’s Article 50 transparency rules, requiring disclosure and labeling of AI-generated content, take effect in August 2026, with penalties reaching €35 million or 7% of global turnover for noncompliance. Meanwhile, roughly 46 to 47 US states have now passed some form of deepfake-specific legislation, spanning election-related disclosure rules, non-consensual imagery protections, and fraud statutes, according to MultiState’s legislative tracking.
None of this stops a scam call from reaching a small business tomorrow morning. But it does signal that lawmakers on both sides of the Atlantic have stopped treating deepfakes as a novelty problem.
Reader Beware: Not Every Stat Holds Up
Scroll through enough 2026 deepfake coverage and you’ll hit percentage increases that sound apocalyptic: 2,137%, 3,892%, four-digit growth claims stacked one after another. A research team at Digital Applied spent its July 2026 audit picking these apart, arguing that the field is crowded with numbers nobody actually verifies, loss figures with no traceable primary source, surge percentages that contradict each other depending on which vendor published them, and forecasts that get recycled as if they were measurements.
Our read: most of those huge percentage jumps are real in direction but misleading in scale. A fraud category that goes from 0.1% to 6.5% of total fraud attempts, which is roughly what’s happened according to fraud-detection firm Signicat, produces an enormous percentage increase almost automatically, simply because it started near zero. That’s still a genuine and fast-growing threat. It’s just not the same thing as the flat “up 3,892% this year” headline that gets repeated without context.
It’s also worth being honest about scale. Most of the largest documented deepfake losses, Arup’s $25.6 million among them, hit large enterprises with the kind of finance operations that can move eight figures in a single transfer. A small business physically can’t lose that much in one incident. The realistic SMB exposure looks more like tens of thousands of dollars per event, which is still enough to close a business operating on thin margins, but the “small businesses are next in line for a $25 million loss” framing overstates the individual stakes even while understating how often SMBs get hit.
The One Habit That Beats the Software
Security researchers keep landing on the same conclusion, and it isn’t a product pitch. Verizon’s Data Breach Investigations Report, cited across multiple 2026 industry analyses, consistently finds the human element involved in more than 60% of breaches. A basic callback-verification habit defeats a deepfake exactly as well as it defeats a decades-old phone scam, because the fake voice or face is only dangerous if the person on the other end skips the second check.
Set a callback rule. Any request to move money, change banking details, or reset credentials gets verified by calling a number pulled from your own records, never one supplied in the suspicious message or call.
Agree on a code word. A pre-shared phrase for high-stakes requests costs nothing and a real-time deepfake can’t guess it.
Slow down on urgency. Scammers manufacture time pressure because it stops people from verifying. Treat “this has to happen right now” as the red flag it is.
Train the one person who approves payments. If your business doesn’t have a finance team, whoever signs off on transfers is your entire defense layer. Make sure they know this playbook exists.
Gartner had already predicted where this was heading: by 2026, the firm projected that 40% of enterprises would stop trusting standalone identity verification because of deepfakes. That prediction is landing now, and the fix it points to isn’t more software, it’s a second channel that a synthetic voice or face can’t fake its way through.
Frequently Asked Questions
What percentage of small businesses have experienced a deepfake scam?
According to VikingCloud’s 2026 SMB Threat Landscape Report, 29% of small businesses reported experiencing a deepfake scheme in the past 12 months, making it one of the most common cyber incidents SMB owners now report, alongside data breaches and ransomware.
How much money has been lost to deepfake and AI-enabled fraud in 2025?
The FBI’s Internet Crime Complaint Center logged $893,346,472 in adjusted losses from 22,364 US complaints referencing AI in 2025, the first year the FBI tracked AI-enabled fraud as its own standalone category.
How can a small business protect itself from deepfake scams?
Require a second-channel verification, a callback to an internally stored phone number or a pre-agreed code word, for any request involving wire transfers, banking-detail changes, or credential resets, even ones that arrive by video call. It consistently ranks above detection software as the lowest-cost, most effective defense.
Why are small businesses targeted by deepfake scammers more than large companies?
Small businesses often rely on informal, trust-based approval processes with no dedicated IT or security staff. Eighty-four percent of SMB owners self-manage their own cybersecurity, per VikingCloud’s 2026 report, which removes the layered sign-off chain that would otherwise catch a fraudulent request.
Can humans reliably spot a deepfake video?
No. Controlled studies find human accuracy at identifying high-quality deepfake videos is only about 24.5%, even though roughly 60% of people believe they could spot one, a gap that itself increases risk by creating false confidence.
Where This Goes Next
Two things are converging right now that weren’t true even a year ago. The FBI has an audited number to point to for the first time, and small business owners are, for the first time in this survey series, ranking cyberattacks above the economy as their biggest worry. Neither of those happens without the other. Watch three things over the next six to eighteen months: whether EU AI Act enforcement actually produces fines large enough to change vendor behavior, whether cyber insurers start pricing deepfake-specific BEC into small business premiums, and whether the “29%” figure gets replicated by a source willing to publish full methodology.
The takeaway for anyone running a small business isn’t to panic about AI. It’s to put a five-minute verification habit in place before you need it. The businesses in the Arup and WPP stories both had smart people on the call. Only one of them had a process that didn’t depend on trusting what they saw.
Nearly 3 in 10 Small Businesses Hit by Deepfake Scams in 2026
NeuralWired Cybersecurity Desk · Published July 23, 2026
In February 2024, a finance employee at UK engineering firm Arup joined what looked like a routine video call with the CFO and several colleagues. He wired $25.6 million across 15 transactions before anyone realized every face on that call except his own was AI generated. Two years later, that trick has trickled all the way down to businesses with a dozen employees and no IT department: 29% of small businesses now say they’ve experienced a deepfake scam in the past year, according to a new survey from cybersecurity firm VikingCloud.
That number, buried inside VikingCloud’s 2026 SMB Threat Landscape Report, is the clearest signal yet that deepfake fraud stopped being an enterprise problem sometime in the last eighteen months. It’s now a Tuesday-afternoon problem for a plumbing company in Ohio or a marketing agency in Manchester. And the FBI, for the first time in its Internet Crime Complaint Center’s roughly 25-year history, agrees the threat is big enough to track on its own.
VikingCloud surveyed small business owners and operators for its 2026 threat report, and the results reorder what SMBs are worried about. More than a quarter said they’d experienced a deepfake scheme (29%), a customer data breach (27%), a ransomware attack (26%), or a denial of service attack (26%) in the past year. Taken together, 75% of SMB owners now rank cyberattacks as their number one operational threat for 2026, the first time in this survey series that cybersecurity has outranked economic pressure. Forty percent said a cyberattack costing $100,000 or less could put them out of business entirely.
A note on the source
VikingCloud hasn’t published full survey methodology, sample size, or margin of error in its public summary; the underlying data sits behind a lead-gen form. That doesn’t make the 29% figure false, but it means it should be read as “according to a vendor survey of small business owners,” not as census-grade data. Compare it against the FBI figure below, which is independently audited.
The FBI Just Made It Official
For 2025, the FBI’s Internet Crime Complaint Center broke out AI-enabled fraud as its own standalone category for the first time. IC3 logged 22,364 complaints with a reported AI nexus, totaling $893,346,472 in adjusted losses, according to the FBI IC3 2025 Annual Report published in April 2026. That figure is the closest thing this space has to a government-audited number, and it’s worth breaking down by category.
Fraud category (AI referenced)
2025 adjusted losses
Investment fraud
$632.0 million
Business email compromise
$30.3 million
Tech and customer-support scams
$19.5 million
Confidence and romance scams
$19.0 million
Employment scams
$12.6 million
Business email compromise is the line that should matter most to a small business owner. It’s the category built entirely around impersonating someone the victim already trusts, a vendor, a boss, a bank contact, and it’s exactly the mechanism behind the Arup case.
The Case That Changed Everything: Arup’s $25.6 Million Call
Arup’s Hong Kong finance team received what appeared to be a standard request from the company’s UK-based CFO: move funds for a confidential transaction. The employee had doubts, so he did what security training tells you to do. He joined a video call to verify. Every other participant on that call, including the person who looked and sounded like the CFO, was an AI-generated deepfake. He made the transfers. Reporting from the Financial Times and CNN in May 2024 confirmed the total loss at $25.6 million across 15 wire transactions, and the case has become the reference point every security vendor cites when explaining why video verification alone is no longer enough.
Arup is a global engineering firm with sophisticated finance operations, not a small business. That distinction matters, and we’ll come back to it. But the mechanics of the attack, real-time video and voice synthesis convincing enough to fool someone who was actively trying to verify, work exactly the same way against a five-person accounting team as they did against Arup’s.
When the Defense Works: WPP’s Near Miss
Not every attempt succeeds, and the counter-example is worth knowing. Scammers targeted WPP CEO Mark Read using a cloned voice and a spoofed Microsoft Teams meeting invite, built around a fake WhatsApp account using his public photo, according to an entry in the OECD.AI Incident Database and reporting from Marketing-Interactive. Staff escalated before any money moved. WPP confirmed zero losses.
What stopped it wasn’t detection software. It was a human asking a question the scammer couldn’t answer and refusing to proceed until someone verified through a separate channel. That’s a cheap lesson, and it’s the same one at the center of the advice section below.
Why Small Businesses Are the Easier Target
Here’s the uncomfortable part for small business owners: being small isn’t protection. It’s the opposite. VikingCloud’s data shows 84% of SMB owners self-manage their own cybersecurity, with no dedicated IT or security staff. That means the same person approving a vendor invoice is also the last line of defense against a fraudulent one, with no gatekeeper, no second sign-off, no layered approval chain to slow things down.
An enterprise like Arup still has structural weaknesses attackers can exploit, but it also has finance controls, compliance teams, and escalation paths. A twelve-person business usually has one bookkeeper and a Slack channel. Attackers know which door is easier to walk through.
“We only have like one really good example in the news right now of that organization in Hong Kong that ended up falling for and sending $25 million based on a deepfake audio and video scam, and I think we’re going to see a lot more business email compromise style events because of AI.”
Rachel Tobac, CEO, SocialProof Security · 8th Layer Insights podcast, The Cyber Wire, April 9, 2024
Tobac’s prediction has aged into the current data. The FBI’s BEC-with-AI-nexus figure alone hit $30.3 million in 2025, and that’s before counting the cases that never get formally reported, which fraud researchers generally assume is the majority of them.
“AI-generated media is not just a future risk, it’s a real business threat. We’re seeing executives impersonated, hiring processes compromised, and financial safeguards bypassed with alarming ease.”
Tony Lee, Head of Consulting, Hong Kong & Macau, Trend Micro · Media OutReach Newswire, July 10, 2025
Worth flagging: Lee’s employer, Trend Micro, sells deepfake detection tools, so treat the quote as an informed but interested voice rather than a neutral one.
Can You Trust Your Own Eyes?
Most SMB owners assume they’d notice if something felt off on a call. The data says otherwise. Controlled lab studies compiled by security research firm DeepStrike found human accuracy at spotting high-quality deepfake video sits at just 24.5%, even though roughly 60% of people believe they could identify one. That gap between confidence and competence is arguably the more dangerous number in this whole story.
The technical barrier to producing convincing fakes keeps dropping too. McAfee’s consumer research found a voice clone with about 85% similarity to the original can now be generated from just three seconds of audio, easily pulled from a podcast clip, a local news interview, or a company’s own marketing video.
The Regulatory Clock Is Ticking
Two regulatory shifts land right around this article’s publish date. The EU AI Act’s Article 50 transparency rules, requiring disclosure and labeling of AI-generated content, take effect in August 2026, with penalties reaching €35 million or 7% of global turnover for noncompliance. Meanwhile, roughly 46 to 47 US states have now passed some form of deepfake-specific legislation, spanning election-related disclosure rules, non-consensual imagery protections, and fraud statutes, according to MultiState’s legislative tracking.
None of this stops a scam call from reaching a small business tomorrow morning. But it does signal that lawmakers on both sides of the Atlantic have stopped treating deepfakes as a novelty problem.
Reader Beware: Not Every Stat Holds Up
Scroll through enough 2026 deepfake coverage and you’ll hit percentage increases that sound apocalyptic: 2,137%, 3,892%, four-digit growth claims stacked one after another. A research team at Digital Applied spent its July 2026 audit picking these apart, arguing that the field is crowded with numbers nobody actually verifies, loss figures with no traceable primary source, surge percentages that contradict each other depending on which vendor published them, and forecasts that get recycled as if they were measurements.
Our read: most of those huge percentage jumps are real in direction but misleading in scale. A fraud category that goes from 0.1% to 6.5% of total fraud attempts, which is roughly what’s happened according to fraud-detection firm Signicat, produces an enormous percentage increase almost automatically, simply because it started near zero. That’s still a genuine and fast-growing threat. It’s just not the same thing as the flat “up 3,892% this year” headline that gets repeated without context.
It’s also worth being honest about scale. Most of the largest documented deepfake losses, Arup’s $25.6 million among them, hit large enterprises with the kind of finance operations that can move eight figures in a single transfer. A small business physically can’t lose that much in one incident. The realistic SMB exposure looks more like tens of thousands of dollars per event, which is still enough to close a business operating on thin margins, but the “small businesses are next in line for a $25 million loss” framing overstates the individual stakes even while understating how often SMBs get hit.
The One Habit That Beats the Software
Security researchers keep landing on the same conclusion, and it isn’t a product pitch. Verizon’s Data Breach Investigations Report, cited across multiple 2026 industry analyses, consistently finds the human element involved in more than 60% of breaches. A basic callback-verification habit defeats a deepfake exactly as well as it defeats a decades-old phone scam, because the fake voice or face is only dangerous if the person on the other end skips the second check.
Set a callback rule. Any request to move money, change banking details, or reset credentials gets verified by calling a number pulled from your own records, never one supplied in the suspicious message or call.
Agree on a code word. A pre-shared phrase for high-stakes requests costs nothing and a real-time deepfake can’t guess it.
Slow down on urgency. Scammers manufacture time pressure because it stops people from verifying. Treat “this has to happen right now” as the red flag it is.
Train the one person who approves payments. If your business doesn’t have a finance team, whoever signs off on transfers is your entire defense layer. Make sure they know this playbook exists.
Gartner had already predicted where this was heading: by 2026, the firm projected that 40% of enterprises would stop trusting standalone identity verification because of deepfakes. That prediction is landing now, and the fix it points to isn’t more software, it’s a second channel that a synthetic voice or face can’t fake its way through.
Frequently Asked Questions
What percentage of small businesses have experienced a deepfake scam?
According to VikingCloud’s 2026 SMB Threat Landscape Report, 29% of small businesses reported experiencing a deepfake scheme in the past 12 months, making it one of the most common cyber incidents SMB owners now report, alongside data breaches and ransomware.
How much money has been lost to deepfake and AI-enabled fraud in 2025?
The FBI’s Internet Crime Complaint Center logged $893,346,472 in adjusted losses from 22,364 US complaints referencing AI in 2025, the first year the FBI tracked AI-enabled fraud as its own standalone category.
How can a small business protect itself from deepfake scams?
Require a second-channel verification, a callback to an internally stored phone number or a pre-agreed code word, for any request involving wire transfers, banking-detail changes, or credential resets, even ones that arrive by video call. It consistently ranks above detection software as the lowest-cost, most effective defense.
Why are small businesses targeted by deepfake scammers more than large companies?
Small businesses often rely on informal, trust-based approval processes with no dedicated IT or security staff. Eighty-four percent of SMB owners self-manage their own cybersecurity, per VikingCloud’s 2026 report, which removes the layered sign-off chain that would otherwise catch a fraudulent request.
Can humans reliably spot a deepfake video?
No. Controlled studies find human accuracy at identifying high-quality deepfake videos is only about 24.5%, even though roughly 60% of people believe they could spot one, a gap that itself increases risk by creating false confidence.
Where This Goes Next
Two things are converging right now that weren’t true even a year ago. The FBI has an audited number to point to for the first time, and small business owners are, for the first time in this survey series, ranking cyberattacks above the economy as their biggest worry. Neither of those happens without the other. Watch three things over the next six to eighteen months: whether EU AI Act enforcement actually produces fines large enough to change vendor behavior, whether cyber insurers start pricing deepfake-specific BEC into small business premiums, and whether the “29%” figure gets replicated by a source willing to publish full methodology.
The takeaway for anyone running a small business isn’t to panic about AI. It’s to put a five-minute verification habit in place before you need it. The businesses in the Arup and WPP stories both had smart people on the call. Only one of them had a process that didn’t depend on trusting what they saw.
Big Tech Layoffs 2026: Why AI Capex Explains It All
Big Tech
Big Tech Layoffs 2026: Why AI Capex Explains It All
Microsoft posted its best quarter ever and cut 4,800 jobs in the same three months. Amazon hit a record 13.1% operating margin and eliminated 30,000 corporate roles. Cisco broke its own revenue record and announced 4,000 layoffs the same week. None of that is a coincidence, and none of it is really about saving money on payroll either. It’s about where the money is actually going.
Big tech layoffs in 2026 keep landing next to record earnings, and the pattern only makes sense once you put the two numbers side by side: what these companies are cutting from headcount, and what they’re pouring into AI infrastructure. The gap between those numbers is the story.
Start with Meta, because the comparison is cleanest there and it sets the pattern for everyone else. Meta’s 2026 capital expenditure guidance sits at $125 billion to $145 billion. Its entire human compensation bill, salaries, benefits, equity, all of it, runs around $27 billion. Even if Meta fired every single employee tomorrow, it wouldn’t cover a fifth of what it’s already committed to spend on AI infrastructure. That comparison comes from a Yahoo Finance analysis of company disclosures, and it’s the single most useful number in this entire story.
Apply the same logic to Microsoft, Amazon, and Cisco and the picture holds. These aren’t companies trimming staff to fund a data center. They’re companies redirecting capital toward compute at a scale where headcount decisions barely register on the balance sheet.
Why this matters: If layoffs were really about cost savings, the numbers would be close. They’re not. Payroll cuts save these companies low single-digit billions. AI capex commitments run into the hundreds of billions. Two completely different orders of magnitude, decided by two largely separate processes inside the same company.
Company
2026 AI Capex
Jobs Cut
Same-Quarter Result
Microsoft
~$190B (guided)
4,800 (plus 9,100 prior round)
Record $82.89B quarterly revenue
Amazon
~$200B (guided)
~30,000 corporate roles
Record 13.1% operating margin
Cisco
$9B AI orders (raised guidance)
Fewer than 4,000 (~5%)
Record $15.84B quarterly revenue
Sources: Microsoft and Amazon Q1/Q3 2026 earnings disclosures; Cisco Q3 FY2026 earnings call.
Microsoft: record revenue, Xbox gutted
Microsoft’s fiscal Q3 2026 numbers, reported April 29, were about as strong as a quarter gets: $82.89 billion in revenue, up 18% year over year, with net income jumping to $31.78 billion from $25.82 billion a year earlier. The company also guided full calendar-year 2026 capex to roughly $190 billion, a 61% jump from 2025 and well past what Wall Street had modeled.
The same quarter, Microsoft cut 4,800 jobs, most of them in the Xbox gaming division, on top of 9,100 roles eliminated about a year earlier. Chief people officer Amy Coleman told staff the cuts weren’t direct AI replacements, even while acknowledging AI is reshaping how the company runs. Microsoft also rolled out its first-ever voluntary buyout program, open to senior director level and below with enough age plus tenure to qualify.
Here’s the part that undercuts the simplest version of the story: Xbox isn’t where Microsoft’s AI money is going. The division that got hit hardest wasn’t competing for capex dollars with Azure’s AI buildout in any direct sense. It just wasn’t the priority, and priority is what actually decides who keeps their job in 2026, not whether AI can technically do the work.
Amazon: 30,000 gone, $200 billion committed
Amazon’s Q1 2026 results, also reported April 29, delivered $181.5 billion in revenue and a record 13.1% operating margin, the highest in the company’s history. AWS grew 28% year over year to $37.6 billion, its fastest growth rate in 15 quarters. Amazon reiterated guidance toward roughly $200 billion in full-year 2026 capex.
Against that backdrop, Amazon cut around 30,000 corporate jobs across rounds in October 2025 and January 2026, with further cuts hitting Selling Partner Services staff and a temporary Homestead, Florida warehouse closure eliminating 600-plus jobs between July and September.
Unlike Microsoft and Cisco, Amazon’s leadership hasn’t tried to soften the connection. CEO Andy Jassy told staff in a memo, later reiterated into 2026, that generative AI and agents would reduce the company’s total corporate workforce over time as efficiency gains materialize, alongside creating new roles elsewhere. That memo dates to July 2025, a full year before this round of cuts, which makes it less a same-day justification and more a stated multi-year strategy Amazon is now executing on schedule.
Cisco: “not a savings-driven restructure”
Cisco reported record quarterly revenue of $15.84 billion on May 13, up 12% year over year, alongside AI infrastructure orders of $2.1 billion that quarter and $5.3 billion cumulative through three quarters. That pushed Cisco to raise its full-year AI order guidance from $5 billion to $9 billion, roughly 4.5 times fiscal 2025’s total.
The same week, Cisco began notifying employees that it would cut fewer than 4,000 jobs, about 5% of its global headcount, with restructuring costs running as high as $1 billion, mostly severance. CFO Mark Patterson gave analysts a line worth sitting with:
“This was really not a savings-driven restructure.”
Mark Patterson, CFO, Cisco Systems, Q3 FY2026 earnings call, via Yahoo Finance
Patterson framed the cuts as a realignment toward silicon, optics, security, and AI rather than a cost play. That’s a notably different posture from Amazon’s Jassy, and it matters: two companies profiled in the same story, cutting staff in the same season, and disagreeing with each other about whether AI is even the reason.
Is AI actually the reason, or the excuse?
Not everyone buys the AI-driven narrative, and the skepticism comes from serious places.
Layoffs are often just standard cost-cutting with an AI label attached.
Paraphrased position of Justin Wolfers, Professor of Economics and Public Policy, University of Michigan, via Benzinga/Finviz
Wolfers argues AI functions as a convenient cover story for restructuring that companies would likely have pursued regardless. JPMorgan’s 2026 economic outlook backs that skepticism with data: the bank’s own labor-market analysis found the AI capex surge hasn’t shown much measurable impact on broader labor dynamics, despite the headlines.
Wall Street’s bull case sees it differently. Wedbush’s Dan Ives, writing about Meta’s own 8,000-role cut against $135 billion in AI capex, called the layoffs financially minor next to the infrastructure commitment, damaging to morale but not decisive to the balance sheet. Evercore ISI’s Mark Mahaney goes further, noting this pattern of workforce actions followed by 12 to 18 months of margin expansion has repeated roughly every one to two years since 2022. In his read, this isn’t new behavior. It’s a recurring capital-discipline cycle that happens to be colliding with an AI narrative people want to believe.
Our read: both things are probably true at once. AI capex is real, historically large, and reshaping where investment goes inside these companies. But the specific decision to cut a specific team often has more to do with which function sits outside this year’s priority list than with any AI system actually replacing a job. Xbox wasn’t cut because a model can ship games. It was cut because it wasn’t where the $190 billion was going.
Worth flagging: Tracking firms don’t agree on the scale of 2026’s layoff wave. Layoffs.fyi-based counts put tech layoffs past 100,000 by early May and over 165,000 by July. SkillSyncer’s broader tracker counts 205,832 people affected across 322 events as of July 22, with 54% of those events explicitly citing AI or automation. The methodologies differ (tech-only versus all-industry, corporate versus contractor), so treat any single total as one tracker’s view, not a consensus figure.
What this means if you work there
If you’re an engineer or manager inside one of these companies, or one like them, the record-revenue headline is not protection. Internal budget decisions are increasingly decoupled from how well your specific team is performing. A well-run division can still get cut if it sits outside whatever core the company is funding this year, silicon, optics, security, and AI at Cisco; cloud and AI infrastructure at Microsoft and Amazon.
The more useful signal than “is the company doing well” is “where is the capex actually going.” Read the earnings call transcript, not just the headline. That’s where you find out whether your function is this year’s priority or this year’s line item.
There’s a real hedge here too. Industry-wide, roughly 275,000 AI-related roles are sitting open while laid-off tech workers largely can’t cross the skills gap to fill them, according to an Invezz analysis of labor-market data. Internal mobility toward AI or infrastructure teams is a legitimate near-term move, but it requires demonstrable fluency, not just tenure at the company.
If you’re evaluating these companies as a vendor rather than an employer, the same logic applies from the other side. Support and account-management staff, the exact functions cut at Amazon’s Selling Partner Services, may thin even as infrastructure capacity grows. That’s worth a line in any vendor risk review.
Frequently asked questions
Why are Microsoft, Amazon, and Cisco laying off workers if their revenue is at record highs?
Revenue and layoffs aren’t directly linked. Each company is redirecting tens of billions toward AI infrastructure capex, roughly $190 billion at Microsoft and $200 billion at Amazon for 2026, while separately restructuring specific divisions that sit outside their AI and cloud growth priorities.
How much is Big Tech spending on AI infrastructure in 2026?
Google, Amazon, Meta, and Microsoft combined are projected to spend roughly $725 billion on AI capital expenditure in 2026, up about 77% year over year, according to aggregated company guidance.
Is AI actually causing the 2026 tech layoffs?
It’s contested. Economists including Justin Wolfers argue AI often serves as a convenient explanation for ordinary cost-cutting. JPMorgan’s own analysis found little measurable labor-market impact from the AI capex surge, even as companies cite AI in layoff announcements.
Did Cisco lay off workers despite good earnings?
Yes. Cisco posted record Q3 FY2026 revenue of $15.84 billion, up 12% year over year, and in the same week announced plans to cut nearly 4,000 jobs as part of a restructuring its CFO described as not savings-driven.
How many tech layoffs have there been in 2026?
Estimates vary by tracker. Layoffs.fyi-based counts show over 165,000 tech layoffs by July 2026, while SkillSyncer’s broader tracker counts 205,832 people affected across 322 events as of July 22, with 54% of events citing AI as a factor.
Where this goes next
What’s changed by walking through all three companies together is this: the “AI is taking jobs” framing is too simple, and so is “it’s just normal cost-cutting.” What’s actually happening is a capital reallocation on a scale large enough that headcount decisions have become almost a separate conversation from infrastructure decisions, loosely connected at best, openly denied at Cisco, openly claimed at Amazon.
Watch three things over the next 6 to 18 months. First, whether Microsoft’s own admission that it will remain capacity-constrained through 2026 even after this spending turns into visible AI revenue, or into a monetization lag that makes the capex look premature. Second, whether more executives start talking like Jassy (AI explicitly reducing headcount) instead of like Patterson (AI reorganizing headcount). Third, whether policymakers, following California’s move this June to build a state tracking tool for AI’s workforce impact, start requiring the kind of capex-versus-headcount disclosure that would make stories like this one unnecessary.
None of these companies are lying when they post record revenue. None of them are lying when they cite AI in a restructuring memo either. They’re just optimizing for two different things at once, and reading the earnings call is currently the only way to tell which one is driving a specific decision.
Want the next capex disclosure and layoff filing broken down like this one? Subscribe to The Neural Loop at neuralwired.com/newsletter.
Your engineering team just spent $4.82 running Claude Opus 4.8 on a routine bug fix that a $0.07 model would have solved just as well. That’s not a hypothetical. It’s the real spread Artificial Analysis measured on its Coding Agent Index this year, and it’s the single most important fact in the best AI models for agentic coding tasks 2026 conversation right now. Model choice used to be about which one scored highest. In 2026, it’s about which one earns its price on the specific task in front of you.
That shift didn’t happen quietly. Six weeks ago, one of the most capable coding models on the market vanished overnight because of a U.S. export control order, then came back three weeks later. Vendors quietly stopped reporting the benchmark everyone used to trust. And developers, according to a JetBrains-backed survey, now spend more hours reviewing AI-written code than writing it themselves. This piece walks through what’s actually true, what’s marketing, and which model belongs on which job.
Why the old benchmarks stopped telling the truth
For most of 2025, SWE-bench Verified was the number everyone quoted. Scores climbed from single digits to the high 80s and low 90s in under two years, a curve that looked like genuine progress until you asked the obvious question: how do models keep getting smarter at solving GitHub issues that were published years before their training cutoff?
In February 2026, OpenAI’s own Frontier Evals team answered that question by walking away from the benchmark entirely. Their reasoning was blunt: model training had absorbed enough of the dataset that the score stopped measuring skill on unseen code and started measuring memorization. An independent audit of the top 30 leaderboard entries found that roughly 19.78% of cases labeled “solved” were passing unit tests by coincidence or by gaming the evaluation harness rather than by producing correct code.
That’s why serious 2026 comparisons have moved to two newer references: SWE-bench Pro, built on private, professional repositories that no model has seen in training, and Terminal-Bench 2.1, which scores the model and its coding harness together as they complete a real terminal-driven task from start to finish. If a vendor is still leading its marketing with a SWE-bench Verified score above 90%, read it the way you’d read a car’s mileage sticker before the EPA got involved.
The 2026 lineup, ranked
Here’s where the six models actually land once you strip out the marketing and look at SWE-bench Pro and Terminal-Bench 2.1, the two benchmarks least contaminated by memorization.
Model
Vendor
SWE-bench Pro
Terminal-Bench 2.1
Pricing (input/output per MTok)
GPT-5.6 “Sol”
OpenAI
Not separately reported
88.8% (highest recorded)
Not disclosed at review time
Claude Fable 5
Anthropic
80.3% (leader)
83.1% (Claude Code)
$10 / $50
Claude Opus 4.8
Anthropic
69.2%
78.9% (Claude Code)
$5 / $25
GPT-5.5
OpenAI
58.6%
83.4% (with Codex)
Not disclosed at review time
Gemini 3.5 Flash
Google DeepMind
55.1%
76.2%
Not disclosed at review time
Grok 4.5
xAI
Not separately reported
Not separately reported
$2 / $6
Two things jump out. First, Claude Fable 5 leads the harder, contamination-resistant benchmark by a wide margin, 11 points ahead of Anthropic’s own Opus 4.8. Second, GPT-5.6 Sol leads the benchmark that best reflects how a coding agent behaves in an actual terminal, doing real multi-step work rather than generating a single patch. Neither model is the “best” one. They’re the best at different jobs.
“But the improvement I keep coming back to is honesty.”
Rahul Patil, CTO, Anthropic, on Claude Opus 4.8’s jump on SWE-bench Pro, via EdTech Innovation Hub
Patil described the target workload for Opus 4.8 as the kind of job that “used to take a quarter and a working group,” meaning codebase-scale migrations and bug fixes spread across hundreds of files. That framing matters. It’s a tacit admission that raw benchmark points matter less than whether the model can survive a genuinely large, messy, real-world job without losing the thread.
Where the open-weight tier fits in
Not every team needs frontier pricing. GLM-5.2 from Z.ai, released under an MIT license, scores 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1 at $1.40/$4.40 per million tokens, the strongest published open-weight coding numbers available right now. Kimi K2.6 Code lands an 80.2% SWE-bench Verified score at open-model pricing, close to Opus-class accuracy for a fraction of the bill. Neither will win a head-to-head against Fable 5 on the hardest tasks. Both will handle the routine 80% of your ticket queue for pennies.
What each task actually costs you
Here’s the number that should reshape how your team budgets for AI coding tools: on Artificial Analysis’s Coding Agent Index, an identical fixed task costs about $0.07 to run through Cursor’s Composer 2.5 and $4.82 to run through GPT-5.5, a roughly 60x spread for only three to four points of quality difference on the index.
The mechanism most teams miss
Coding agents burn most of their budget on reading, not writing. PointFive’s July 2026 index found that a single realistic task, reading a handful of files, reasoning about them, and revising a diff, can pull in up to 200,000 input tokens against a diff of roughly 30,000 tokens written back. That’s why input pricing, and whether a model caches repeated reads at a discount (often around 10% of the standard rate), swings your real bill far more than the headline output price per token.
Zoom out and the frontier-to-workhorse spread gets even wider. Claude Fable 5 charges $10/$50 per million tokens. DeepSeek V4 Flash charges roughly $0.14. That’s close to a 180x difference in raw token pricing between the most expensive and cheapest options a team might reasonably put in production this year.
None of this means cheap wins by default. GLM-4.6 costs about $0.059 per task but is statistically tied on accuracy with pricier open options, which means the math sometimes favors a marginally more expensive model like DeepSeek or Qwen instead. The lesson isn’t “buy the cheapest model.” It’s “stop assuming the most expensive model is the safest default,” and start routing tasks by difficulty: cheap model first, escalate to a frontier model only when the cheap one fails.
The Fable 5 warning every team should have caught
Claude Fable 5 launched June 9, 2026, and immediately topped the SWE-bench Pro leaderboard. Three days later, on June 12, 2026, Anthropic suspended it worldwide to comply with a U.S. Department of Commerce export control order. Access came back on July 1, 2026, after the controls were lifted, and Anthropic confirmed the restoration directly.
Three weeks of downtime for a model teams were actively shipping in production. If your pipeline depended entirely on Fable 5 during that window, you didn’t have a benchmark problem. You had a supply chain problem, and most engineering leaders still aren’t tracking it as one.
There’s a second wrinkle independent evaluators caught after Fable 5 came back online: Artificial Analysis and Vals AI both measured Fable 5 refusing roughly 8 to 9% of test prompts, quietly falling back to Opus 4.8 for those cases. That means the headline SWE-bench Pro score doesn’t fully describe what a production deployment experiences. A meaningful slice of real traffic never actually touches the model you thought you were paying for.
Best practice going forward: never build a single-model dependency into a critical pipeline. Keep at least one fallback model configured, and treat a vendor’s top model the way you’d treat a single-region cloud deployment. It works great, right up until it doesn’t.
The bottleneck nobody’s marketing deck mentions
Every vendor above is racing to add benchmark points. Almost none of them are talking about the actual reason enterprise AI coding adoption stalls, and that’s reliability, not raw capability.
“It unpacks different factors that I see tangled together in almost every eval I’ve ever seen.”
Bryan Silverthorn, Director of AGI Autonomy, Amazon, at VB Transform 2026
Silverthorn, who joined Amazon through its Adept AI acquisition, argues that “reliability” isn’t one thing. He breaks it into four separate dimensions, borrowing a framework from Princeton research: consistency, robustness, predictability, and safety. He described a customer whose agent performed a serial number extraction task flawlessly for two months, then quietly started misreading numbers with no warning and no obvious trigger. No benchmark on this list would have caught that failure mode before it hit production.
Paul Gauthier, creator of the open source pair-programming tool Aider, has built a reputation on the opposite end of the spectrum: refusing to rank his own tool against agents that won’t publish their evaluation methodology. If a vendor won’t show its work, that’s a signal worth weighing as heavily as the score itself.
There’s a human cost showing up in the data too. A developer survey compiling adoption research found that engineers using AI coding tools now spend 11.4 hours a week reviewing AI-generated code, against 9.8 hours writing new code themselves, a reversal from the pattern two years ago. The “10x productivity” pitch quietly assumes review time is free. It isn’t.
How to actually choose, task by task
Stop asking which model is smartest. Ask which model fits the task type sitting in your queue right now.
CLI-heavy DevOps and multi-step terminal work: GPT-5.6 Sol currently leads Terminal-Bench 2.1 at 88.8%, the strongest publicly reported score for real terminal-agent workflows.
The hardest multi-file repository repairs: Claude Fable 5 leads SWE-bench Pro, provided you’ve built in a fallback for its 8 to 9% refusal rate and you’re comfortable with the export control volatility above.
Large-scale migrations and refactors across hundreds of files: Claude Opus 4.8, purpose-built by Anthropic for exactly this workload, with its Dynamic Workflows feature fanning work out to parallel subagents.
Tool-orchestration-heavy agent work: Gemini 3.5 Flash leads MCP Atlas at 83.6% even though it trails on raw SWE-bench numbers, making it a genuine specialist pick for agentic tool-calling.
The routine 80% of your ticket queue: An open-weight model like GLM-5.2 or Kimi K2.6, or a workhorse like Cursor’s Composer 2.5, saves 10 to 60x on cost for a 3 to 4 point accuracy trade-off most teams won’t even notice.
Our read: the real 2026 skill isn’t picking a single model and standardizing on it. It’s building a routing layer that sends each task to the cheapest model likely to solve it, and escalates only on failure. Teams still budgeting per seat instead of per completed task are leaving real money on the table, and the PointFive and Artificial Analysis data above shows exactly how much.
Frequently asked questions
What is the best AI model for coding in 2026?
There’s no single winner. Claude Fable 5 leads the hardest contamination-resistant benchmark, SWE-bench Pro. GPT-5.6 Sol leads real terminal-agent work, scoring 88.8% on Terminal-Bench 2.1. The right choice depends on task type and budget, and open-weight models like GLM-5.2 close most of the gap at a fraction of the cost.
How much does an AI coding agent cost per task?
Cost per completed coding task ranges from roughly $0.07 to $4.82 depending on the model, according to Artificial Analysis and PointFive benchmark data. Workhorse models like Cursor’s Composer 2.5 cost around $0.07 per task, while frontier models like GPT-5.5 or Claude Opus can run $4 or more for only a few extra benchmark points.
Why did OpenAI stop reporting SWE-bench Verified scores?
OpenAI’s Frontier Evals team announced in February 2026 that it would stop reporting SWE-bench Verified results because training data contamination had inflated scores past the point where they reflected real coding ability on unseen code. SWE-bench Pro, built on private repositories, is now the more trusted reference.
Is Claude Fable 5 still available?
Yes. Claude Fable 5 launched June 9, 2026, was suspended worldwide on June 12, 2026 under a U.S. Department of Commerce export control order, and access was restored on July 1, 2026 after the controls were lifted. Teams building on it should keep a fallback model plan in place given that volatility.
Where this goes next
The benchmark story of 2026 is really a trust story. Vendors spent two years optimizing for a number that eventually stopped meaning anything, and the market is only now rebuilding around harder, more honest measures like SWE-bench Pro and Terminal-Bench 2.1. Cost-per-task, not leaderboard rank, is fast becoming the metric that actually determines what ships to production.
Three things worth watching over the next six to eighteen months: whether Anthropic can keep Fable 5 and Mythos 5 available without another export control disruption, whether the 60x cost gap between frontier and workhorse models narrows as competition in the open-weight tier intensifies, and whether reliability metrics like Bryan Silverthorn’s four-part framework get standardized into a benchmark of their own. Gartner’s projection that 40% of new enterprise production software will involve vibe coding by 2028 is a forecast, not a fact on the ground today, and it deserves the same skepticism this piece just applied to SWE-bench Verified.
Want the next model launch, export control ruling, and cost benchmark broken down the same way? Subscribe to The Neural Loop at neuralwired.com/newsletter.
Best LLM Inference Optimization Tools 2026: 7 Engines Tested
Infrastructure / Developer Tools
vLLM vs SGLang vs TensorRT-LLM: 7 Inference Engines Tested Against MLPerf v6.0
By the NeuralWired Infrastructure Desk | Published July 22, 2026 | 12 min read
Your GPU bill went up again last month, and your throughput barely moved. That’s not a hardware problem anymore. It’s a software problem, and in 2026 the gap between a well-tuned inference stack and a default install is wide enough to change your infrastructure budget by double digits.
This guide ranks the best LLM inference optimization tools of 2026 for teams actually running models in production, not just testing them on a laptop. We pulled numbers from MLCommons’ newly released MLPerf Inference v6.0 suite, cross checked vendor claims against SemiAnalysis’s InferenceX benchmark platform, and flagged where the industry’s own benchmarking methods might be lying to you.
Two years ago, “serving an LLM” meant getting Llama 2 to answer chat prompts fast enough that users didn’t notice the wait. In 2026, the traffic looks nothing like that. Reasoning models like DeepSeek-R1 chew through multi-step chains of thought before producing a token. Agentic workloads fire off dozens of overlapping requests with shared context. RAG pipelines lean hard on prefix reuse. None of that fits the old single-turn chat benchmark.
That shift is exactly what MLCommons built its latest benchmark suite to capture, and it’s why picking an inference engine in 2026 is a genuinely different decision than it was in 2024.
What MLPerf Inference v6.0 Actually Measured
MLCommons released MLPerf Inference v6.0 on April 1, 2026, and it’s the most substantial rewrite of the suite in the benchmark’s history. Five of eleven datacenter tests were new or updated: a GPT-OSS 120B benchmark for math and coding reasoning, an expanded DeepSeek-R1 test with a speculative decoding scenario, a Meta-contributed recommender benchmark called DLRMv3, the suite’s first text-to-video generation test, and a new vision-language benchmark built on Shopify product catalog data.
“This is the most significant revision of the Inference benchmark suite that we’ve ever done.”
Frank Han, Systems Development Engineering, Dell Technologies, MLPerf Inference Working Group Co-chair
The suite also introduced a new harness called LoadGen++, which lets submitters run benchmarks against a serving style software stack that looks a lot more like real production traffic than the older synthetic load generators. Twenty four organizations submitted results this round, including AMD, Google, NVIDIA, Oracle, Red Hat, and Lambda.
The infrastructure scale tells its own story. Multi-node submissions jumped 30% compared to the prior round in September 2025, and the largest submitted system used 72 nodes and 288 accelerators, four times the node count of the previous record. Ten percent of submitted systems now use more than ten nodes, up from just 2%.
“These partnerships were essential in ensuring that the tests include scenarios and workloads that represent the current state of the industry.”
Miro Hodak, Senior Member of Technical Staff, AMD, MLPerf Inference Working Group Co-chair
Ultralytics contributed an upgraded YOLOv11 based object detection test to the edge category. Its founder framed the value of the whole exercise plainly.
“MLPerf Inference benchmarks play a vital role in driving transparency and accountability across the AI industry.”
Glenn Jocher, CEO and Founder, Ultralytics
Why this matters for your engine choice
If you’re still benchmarking candidate engines against single-turn chat throughput, you’re testing for a workload that’s disappearing. MLPerf’s own suite has moved to reasoning and multi-node scenarios because that’s where production traffic actually lives now.
The 7 Engines, Compared
Here’s where most “best of” lists go wrong: they rank engines on raw tokens per second and call it a day. Workload shape matters more than any single number. A RAG pipeline and a batch summarization job want different things from the same GPU.
Engine
Core approach
Best fit
Notable stat
vLLM
PagedAttention memory management, broad hardware support
General purpose, fastest path to production, multi vendor hardware
~86,000 GitHub stars, releases roughly every two weeks
SGLang
RadixAttention prefix cache reuse
RAG, multi-turn chat, agentic loops with shared context
Frequently cited as the strongest option for prefix-heavy traffic
TensorRT-LLM
Compiles models into NVIDIA’s proprietary engine format
Current version referenced in 2026 sources: v1.2.0
llama.cpp
C/C++ engine, the backbone under most local tools
CPU and consumer-hardware inference, embedded deployments
121,000 GitHub stars, 1,806 contributors
Ollama
llama.cpp backend (MLX reported on Apple Silicon in 2026)
Local developer workflows, prototyping
Reported at 172,000+ GitHub stars (unverified against GitHub directly)
Hugging Face TGI
Originally general purpose serving toolkit
Legacy deployments only, per reported status below
Reported to be in maintenance mode in 2026, unconfirmed against source repo
InferenceX benchmark stack
Not an engine; a continuously updated benchmark platform
Validating vendor claims before you commit to an engine
V2 launched February 2026 with GB300 NVL72 coverage
A word on that Hugging Face TGI line, because it matters for anyone planning a migration. Multiple 2026 sources describe TGI shifting into maintenance mode with new users pointed toward vLLM, SGLang, and llama.cpp instead. We could not independently confirm this against Hugging Face’s own repository at the time of writing, so treat it as reported rather than settled. If you’re running TGI in production today, check the repository directly before you build a migration plan around a secondhand claim.
Worth noting too: 2026 is the year inference optimization became its own funded category rather than a side effect of the model layer. vLLM’s own creators, Woosuk Kwon, Simon Mo, and Ion Stoica, launched a company called Inferact in January 2026, backed by a16z, specifically to build what they call a universal inference layer across hardware and model architectures. When the people who solved the original memory bottleneck go out and raise venture money to solve it again commercially, that tells you how much money is riding on this layer of the stack.
Software Beats Hardware More Often Than You’d Think
Here’s the number that should reframe how you think about your next GPU purchase: in its own MLPerf v6.0 submission, Lambda found that NVIDIA’s Blackwell Ultra delivered 29% more throughput than the prior Blackwell generation on identical workloads. But the software stack alone, running on the exact same hardware, added another 9%. Lambda’s Smart Expert Routing technique cut P99 time-to-first-token by 31%.
Read that again. A software change on unchanged hardware moved the needle nearly a third as much as an entire GPU generation upgrade. If your team is budgeting for next year’s inference costs purely around which GPUs to buy, you’re solving half the problem.
This tracks with the broader MLPerf v5.1 data from September 2025, where NVIDIA’s GB300 NVL72 (Blackwell Ultra) delivered 45% more DeepSeek-R1 reasoning throughput than the prior GB200 NVL72 generation, according to HPCwire’s coverage of that round. AMD also made its first submissions that cycle, with a 4-node MI355X cluster posting a 3.4x throughput gain over the prior MI300X generation.
The Benchmark Trust Problem Nobody’s Talking About
Here’s our contrarian take, and it’s not ours alone: a lot of the “Engine X beats Engine Y by 40%” claims circulating in 2026 comparison articles might be measuring the wrong thing entirely.
A 2026 preprint from Google researchers Ashok Chandrasekar and Jason Kramberger modeled the client side of common LLM benchmarking tools using queueing theory, and found that single-process, asyncio-driven benchmarking clients suffer from Python’s GIL creating a queuing bottleneck of their own. As concurrency scales up, that bottleneck artificially inflates time-to-first-token and time-per-output-token measurements, according to the paper on arXiv. In plain terms: the tool measuring the engine can be slower than the engine it’s measuring, and that gap gets bigger exactly when you push concurrency higher, which is when it matters most.
Our read
This signals that a meaningful share of the public “vLLM vs SGLang vs TensorRT-LLM” benchmark posts circulating right now may be comparing benchmarking tool artifacts as much as real engine performance. Before you make a purchasing or migration decision off a single blog post, check whether the methodology discloses its process model and concurrency handling. If it doesn’t, treat the numbers as directional at best.
This isn’t a hypothetical concern either. Different sources genuinely disagree on the same matchups. One benchmark shows SGLang and LMDeploy topping 16,000 tokens per second against vLLM’s roughly 12,500 on Llama 3.1 8B, while another shows TensorRT-LLM leading at every concurrency level on Llama 3.3 70B running on H100. These aren’t necessarily contradictory since the model, hardware, and settings all differ, but it does mean a single “best overall” ranking is a simplification you should be skeptical of.
The safer move: validate any vendor benchmark against MLCommons’ public MLPerf dashboard or against SemiAnalysis’s InferenceX platform, which is continuously updated with support from Crusoe, CoreWeave, Nebius, TensorWave, Oracle, and Together AI. Crusoe’s CEO put the case for that kind of open, reproducible testing directly.
“At Crusoe, we believe being a great partner means empowering our customers with choice and clarity. That’s why we’re proud to support InferenceMAX, which provides the entire AI community with open source, reproducible benchmarks for the latest hardware.”
Chase Lochmiller, Co-Founder and CEO, Crusoe
How to Actually Choose One
Skip the “best overall” instinct. Match the engine to the traffic pattern.
Building a RAG or agentic product with heavy prefix reuse? SGLang’s RadixAttention scheduling is repeatedly cited as the stronger fit because it caches and reuses shared context instead of recomputing it.
Need hardware portability across NVIDIA, AMD, TPUs, or Trainium? vLLM remains the broadest bet, with the fastest path from prototype to production and a release cadence that outpaces almost every other infra project in this category.
All-in on NVIDIA and chasing maximum single-model throughput? TensorRT-LLM’s compiled engine format still wins on raw numbers once you’ve paid the setup cost.
Running locally or on consumer hardware? llama.cpp is the substrate under nearly every local tool, including Ollama, and its 121,000 GitHub stars and 1,806 contributors reflect how entrenched it’s become.
Still on Hugging Face TGI? Confirm its current status directly against the repository before you plan around secondhand reports of maintenance mode.
And one budgeting reality check: per-token inference costs for GPT-4-class performance have reportedly fallen roughly 1,000x over three years, down to around $0.40 per million tokens by some estimates. Total inference spending keeps rising anyway, because usage growth is outpacing the unit cost declines. Switching engines can absolutely cut your cost per token. It won’t necessarily shrink your total bill, and any plan that assumes otherwise is setting up a budget conversation you’ll lose later.
Frequently Asked Questions
What is the best LLM inference engine in 2026?
There’s no single “best” engine, it depends on workload. vLLM offers the broadest hardware support and fastest path to production. SGLang leads on prefix-heavy workloads like RAG and multi-turn chat via RadixAttention. TensorRT-LLM delivers the highest raw throughput on NVIDIA-only deployments once compiled.
What is the difference between vLLM and TensorRT-LLM?
vLLM is an open source, hardware-portable engine (NVIDIA, AMD, TPU, Trainium) using PagedAttention for memory efficiency and fast iteration. TensorRT-LLM is NVIDIA’s proprietary engine that compiles models into optimized TensorRT engines for maximum single-vendor throughput, at the cost of flexibility and longer setup time.
Is Hugging Face TGI still maintained in 2026?
Multiple 2026 sources report TGI has moved into maintenance mode, with Hugging Face directing new projects toward vLLM, SGLang, and llama.cpp instead. This is reported by secondary sources and should be confirmed directly against Hugging Face’s own repository before being treated as final.
What is MLPerf Inference and why does it matter for LLM serving?
MLPerf Inference is MLCommons’ independently audited, vendor-neutral benchmark suite for AI system performance. Its April 2026 v6.0 release added GPT-OSS 120B, expanded DeepSeek-R1 reasoning, and a new serving-style LoadGen++ harness, making it the most credible public reference point for comparing real-world inference performance.
How much does LLM inference cost in 2026?
Estimates vary, but per-token costs for GPT-4-class performance are widely reported to have fallen roughly 1,000x over three years, to around $0.40 per million tokens. Total inference spending continues rising anyway, because usage growth is outpacing these per-token cost declines.
Where This Goes Next
The engine layer stopped being a free add-on to the model layer somewhere around January 2026, when vLLM’s own founders decided it was worth its own venture-backed company. Expect more of that: standalone inference infrastructure businesses raising money on the premise that the serving stack, not the model weights, is where the next round of margin gets won or lost.
Watch three things over the next 6 to 18 months. First, whether Hugging Face TGI’s reported maintenance mode status gets confirmed or walked back, since that will settle a live migration debate for a lot of teams. Second, whether InferenceX and MLPerf’s LoadGen++ harness push more vendors toward disclosing their actual benchmarking methodology, given the Google GIL findings. Third, whether reasoning-model workloads like DeepSeek-R1 and GPT-OSS 120B keep pulling multi-node deployment further into the mainstream, the way the 30% jump in multi-node MLPerf submissions suggests they already are.
Pick your engine for the traffic you actually have, not the traffic you had two years ago. And test any benchmark claim, including the ones in this article, against your own workload before you bet a production budget on it.
By NeuralWired Staff | Updated July 22, 2026 | 9 min read
The crypto market has lost more than half its value since October 2025. Financial crime running through it just hit a record anyway. That is the story the numbers actually tell in July 2026, and it is a more useful one than the stale “$4 trillion milestone” headlines still floating around search results.
On July 16, 2026, the Financial Action Task Force (FATF) published its 7th Targeted Update on Virtual Assets and VASPs, and the headline finding should worry anyone running compliance at a crypto exchange or fintech: organized crime networks are no longer just laundering money through existing stablecoins like USDT. Some are now building their own, engineered specifically so authorities cannot freeze or seize them.
The Market Reality Check Nobody’s Headline Is Running
Start with the number everyone gets wrong right now. The total crypto market cap first crossed $4 trillion on July 17 and 18, 2025, then peaked around $4.27 to $4.35 trillion on October 6, 2025, fueled by a Bitcoin rally past $125,000 during a U.S. government shutdown, according to The Block’s original reporting. Within 24 hours of that peak, a leverage unwind wiped out roughly $19 billion in positions.
It never recovered. By the end of Q2 2026, total market cap had fallen to $2.1 trillion, down 12.6% for the quarter alone and the lowest reading since September 2024, per CoinGecko’s Q2 2026 Crypto Industry Report. Bitcoin itself was trading around $66,544 as of July 21, 2026, less than half its October high.
If your instinct says crime should have shrunk along with the market, the FATF update says otherwise. That gap between falling prices and rising criminal sophistication is the actual story.
Why the “$4 trillion” framing keeps resurfacing
Several 2026 articles reused the “crypto hits $4 trillion for the first time” line as a retrospective reference while covering unrelated news. It is easy to mistake that for current data if you’re skimming. As of July 2026, the market sits at roughly half that figure.
What FATF Actually Found
FATF’s update lands at the start of the UK’s FATF presidency, and the tone is blunt for an intergovernmental body that usually hedges. Giles Thomson, FATF President, put it this way:
“Governments and the private sector must work together to strengthen preventive measures and close regulatory gaps.”
Giles Thomson, President, Financial Action Task Force, July 16, 2026
The report’s most consequential finding is the proprietary stablecoin problem. Criminal networks are moving past simply misusing USDT or USDC. They are issuing their own tokens, built from the ground up to resist the freeze-and-seize tools that make blockchain forensics useful to investigators in the first place.
One case in the report: a Cambodia-based financial services conglomerate laundered at least $4 billion between 2021 and 2025, serving both organized fraud rings and DPRK-linked cyber theft operations through shared infrastructure. Separately, Spain’s Guardia Civil dismantled a crypto investment fraud network in June 2025 that allegedly laundered close to 460 million euros from more than 5,000 victims worldwide.
The Numbers Behind the Warning
Pull the threads together and the scale gets clearer. Here’s what the data shows across the last reporting cycle.
Metric
Figure
Source
Value received by illicit crypto addresses, 2025
$154 billion (+162% YoY)
Chainalysis
Value received by sanctioned entities, 2025
$104 billion (+694% YoY)
Chainalysis
U.S. crypto fraud losses reported to FBI IC3, 2025
$11.37 billion (record high)
FBI Internet Crime Complaint Center
Losses reported by victims aged 60+
$4.4 billion
FBI IC3
Jurisdictions “largely compliant” with FATF standards
34% of 149 assessed
FATF, 7th Targeted Update
That FBI figure deserves a beat of its own. Crypto-related fraud accounted for more than half of the $20.877 billion in total internet crime losses the FBI tracked across all categories in 2025, drawn from 181,565 complaints with an average loss of $62,604 per victim, according to the FBI’s 2025 Internet Crime Report, released April 6, 2026. It was the first year IC3 logged more than one million total complaints since the unit’s founding.
Inside the Seizure-Proof Stablecoin Problem
Here’s the mechanism that makes FATF’s warning concrete rather than theoretical. A7A5, a ruble-backed stablecoin, processed $93.3 billion in transaction volume in under ten months in 2025, functioning as a sanctions-evasion settlement rail. Its “Instant Swapper” service alone moved $2.2 billion into mainstream USD stablecoins with little to no identity verification, according to Chainalysis’s 2026 sanctions research.
Then there’s Huione Group, which processed $98 billion in cryptocurrency inflows between August 2021 and January 2025, including more than $4 billion in confirmed illicit proceeds. It was designated under FinCEN Section 311 in October 2025. Related enforcement action seized $15 billion from Prince Group frontman Chen Zhi, one of the largest single asset seizures tied to crypto-enabled crime on record.
These aren’t edge cases. They’re the working prototype for exactly what FATF flagged this month: purpose-built financial infrastructure designed by criminal organizations, not adapted from legitimate tools after the fact.
The Counterargument: Is This Overblown?
Not every voice in this space agrees the sky is falling, and the skepticism deserves airtime. Chainalysis’s own data shows illicit activity represented just 0.14% of total blockchain transaction volume in 2024. Compare that to the United Nations Office on Drugs and Crime’s estimate that 2 to 5% of global GDP is laundered through the traditional banking system every year, and crypto’s proportional crime rate looks small next to the system it’s often compared against unfavorably.
There’s also a methodology question that regulators lean on more than they acknowledge. In the Roman Sterlingov Bitcoin Fog case, defense expert Jonelle Still, then director of investigations at CipherTrace, called the blockchain clustering methodology used to build the prosecution’s case “unverifiable” in court testimony reported by Fortune in March 2024. No independent standards body audits these attribution tools, yet FATF, FinCEN, and law enforcement worldwide increasingly build policy on top of them.
SEC Commissioner Hester Peirce, who heads the agency’s Crypto Task Force, offered a notably different framing to industry at a Katten Muchin Rosenman symposium this summer:
“This is your moment. You’ve got agencies that want to work with you.”
Hester Peirce, Commissioner, U.S. Securities and Exchange Commission, 2026
That’s the tension worth sitting with. U.S. regulators are simultaneously loosening the rules to encourage innovation while FATF and the FBI report record crime figures. Both things are true at once, and neither cancels the other out.
What This Means for Compliance Teams
If you run compliance at an exchange, custody provider, or fintech touching crypto rails, the proprietary stablecoin finding changes your actual workload, not just your risk register.
Static token blacklists are no longer enough. Screening built around known assets like USDT and USDC misses tokens criminal groups mint themselves specifically to dodge those lists.
Jurisdictional exposure is indirect now. With only 34% of assessed jurisdictions rated largely compliant, a fully compliant exchange can still inherit sanctions risk through a counterparty routed via a weaker jurisdiction.
Crime doesn’t track price. Illicit volume and sophistication rose straight through the 2026 downturn. Budgets tied to the assumption that crime scales down with market cap are working from a false premise.
False positives carry a real cost. Crypto compliance programs already see amplified false-positive rates compared to traditional finance, according to Elliptic’s April 2026 analysis, which means more monitoring alerts do not automatically mean less actual crime caught.
Our read: this signals a shift from reactive blacklisting to behavior-based monitoring as the baseline expectation, not the advanced option, for any VASP operating across borders.
Frequently Asked Questions
What is the current crypto market cap?
As of July 2026, the total crypto market cap is roughly $2.1 to $2.3 trillion, down about 52% from its October 2025 all-time high of $4.27 trillion, according to CoinGecko’s Q2 2026 report.
When did crypto market cap hit $4 trillion?
The global crypto market first crossed $4 trillion on July 17 and 18, 2025, then peaked near $4.27 to $4.35 trillion on October 6, 2025, before entering a prolonged 2026 downturn.
How much crypto crime was there in 2025?
Chainalysis found illicit crypto addresses received at least $154 billion in 2025, up 162% year over year, driven mainly by a 694% surge in sanctions-evasion volume. Separately, the FBI logged $11.4 billion in crypto fraud losses from U.S. victims alone.
What did the FATF report on crypto say in 2026?
FATF’s July 2026 targeted update found organized crime groups moving billions through crypto, including networks that have built proprietary stablecoins specifically designed to resist freezing or seizure by authorities.
Why are stablecoins linked to crypto crime?
Most identified illicit on-chain activity now involves stablecoins because they offer price stability and liquidity for large-scale settlement. FATF’s 2026 update found criminal groups increasingly issuing their own stablecoins built to resist asset freezing.
Where This Goes Next
Here’s what’s different after this week: the case for treating crypto crime as a function of market size just fell apart. Prices are down 52%. Illicit volume, sanctions evasion, and criminal infrastructure sophistication all moved the opposite direction through the same period.
Watch three things over the next 6 to 18 months. First, whether FinCEN and its European counterparts start designating proprietary criminal stablecoins by name, the way they did with Huione Group in October 2025. Second, whether the 34% “largely compliant” jurisdiction figure moves meaningfully before FATF’s next review, since that number is the real chokepoint on enforcement. Third, whether behavior-based transaction monitoring becomes a stated regulatory requirement rather than a best practice exchanges adopt voluntarily.
The market will keep doing what markets do. The compliance problem isn’t going to wait for it to recover.
Want stories like this before they hit your feed? Subscribe to The Neural Loop at neuralwired.com/newsletter.
64% of Institutions Are Now Tokenizing Assets: Inside the 2026 Enterprise Blockchain Market
Enterprise Blockchain · Market Analysis
64% of Institutions Are Now Tokenizing Assets: Inside the 2026 Enterprise Blockchain Market
JPMorgan’s Kinexys platform just crossed $4 trillion in cumulative volume. The Federal Reserve says tokenized assets doubled in a year. But the “market size” numbers everyone’s citing don’t agree with each other, and one popular stat about institutional adoption is being misquoted across the web.
By The Neural Loop Research Desk · NeuralWired.com · Updated July 20, 2026
Somewhere between a JPMorgan press release and a market-research PDF, a number got mangled. You’ve probably seen it: “67% of institutions are prioritizing tokenization.” It’s been repeated across newsletters, LinkedIn posts, and at least one aggregator site as though it settles the question of how fast the enterprise blockchain market 2026 story is moving.
It doesn’t say that. And chasing the wrong number matters, because the real numbers tell a more interesting story anyway: JPMorgan’s institutional settlement platform just crossed $4 trillion in cumulative volume, the Federal Reserve is on record saying tokenized assets in the U.S. roughly doubled in a single year, and Citi thinks the market could hit $5.5 trillion by 2030. None of that requires an inflated stat to be compelling.
The stat everyone’s getting wrong (and the real number underneath it)
Fact-check: the “67% prioritizing tokenization” claim is a misread.
The 67% figure comes from Coinbase and EY-Parthenon’s 2026 Institutional Investor Digital Assets Survey, a poll of 351 institutional decision-makers run in January 2026. In the original survey, 67% of respondents named regulatory uncertainty as the single biggest barrier to investing in tokenized assets. That’s a stat about hesitation, not enthusiasm. Somewhere in the retelling, “biggest barrier” became “prioritizing,” and the meaning flipped.
Here’s what the same survey actually found on adoption: 64% of asset managers say they’re interested in tokenizing their own assets, up from 40% just a year earlier. That’s a 24-point jump in twelve months, which is a genuinely large swing for an institutional survey, and it didn’t need to be dressed up as something else.
The survey also found 63% of investors are interested in allocating capital to tokenized assets, and 66% now cite regulatory compliance as their top factor when picking a custodian, up from just 25% the year before. Read together, the picture isn’t “institutions are racing in.” It’s “institutions want in, and they’re building compliance infrastructure first.” That’s a slower, more credible story than a viral stat, and it happens to be true.
Enterprise blockchain hit production scale in 2026, not pilot scale
For years, the standard skeptic line on enterprise blockchain was fair: lots of pilots, not much production volume. That line stopped being accurate sometime in the first half of 2026.
Start with JPMorgan. Its blockchain settlement platform, rebranded Kinexys in late 2024, announced an expansion on June 29, 2026 that added five Asia-Pacific currencies (Australian dollar, Hong Kong dollar, Japanese yen, Chinese renminbi, and Singapore dollar) to the three it already supported. That brings Kinexys to eight currencies running round-the-clock cross-border settlement on a permissioned ledger. The volume behind that expansion is the part that should get your attention: more than $4 trillion processed cumulatively, with average daily volume now exceeding $7 billion.
Oliver Harris, who took over as Head of Kinexys in April 2026 after leading digital assets at Goldman Sachs, isn’t running an experiment. He’s running settlement infrastructure that banks route real money through, every day, in eight currencies.
Broadridge tells a similar story from a different corner of the market. Its Distributed Ledger Repo platform, which handles repurchase agreements, not exactly a headline-grabbing product category, processed $8 trillion in March 2026 alone. That’s 392% year-over-year growth, with daily volume exceeding $400 billion. Broadridge is back-office plumbing, not a bank brand chasing press coverage, which makes the growth number harder to dismiss as marketing.
“I think that we’re at an inflection point right now… not should we, but how much?”
Ryan Rugg, Global Head of Digital Assets, Citi Treasury and Trade Solutions · PYMNTS “From the Block” podcast, January 29, 2026
Rugg’s read is worth sitting with, because she’s not a hype account. In the same interview, she was explicit that permissionless DeFi protocols “in their purest form” aren’t getting embraced by regulated institutions any time soon. Her framing for 2026 is re-architecture, not disruption, and she expects adoption to stay “messy and uneven” rather than sweeping.
Morgan Stanley’s CFO struck a similar note of seriousness, minus the caution, on the company’s Q1 2026 earnings call.
“How do you think of a tokenized world? How do you think of an onchain world where you can move assets quickly, the same way you’d be able to move those liabilities quickly?”
Sharon Yeshaya, Chief Financial Officer, Morgan Stanley · Q1 2026 earnings call, reported by CoinDesk, April 15, 2026
That’s a CFO of a multi-trillion-dollar wealth management business framing tokenization as core infrastructure strategy, not a side bet for the innovation team. If you’re wondering whether this topic cleared the “should we care” threshold this year, that’s your answer.
Why “market size” numbers disagree by 20x, and what to actually trust
Here’s where the enterprise blockchain market 2026 conversation gets genuinely messy, and where most coverage quietly skips the hard part.
Pull the “global enterprise blockchain market” figure for 2026 from six different research firms in the same week, and you get numbers that don’t remotely agree.
