Anthropic’s AI Is Improving Twice as Fast as a Year Ago, Independent Audit Finds — With AI Itself Doing Most of the Work

Artificial intelligence is now improving itself at roughly twice the speed it managed a year ago, according to an independent audit published on Thursday — and the finding arrives with a warning attached: the company at the centre of it removed most of the evidence that would let the public check the claim for itself.

The assessment, carried out by the nonprofit research group METR inside Anthropic over eleven weeks, found that the maker of the Claude chatbot is improving its models on well-specified tasks at about twice the rate it achieved between March 2024 and October 2025, according to METR. On the group’s main yardstick — Anthropic’s version of the Epoch Capabilities Index, known as AECI — benchmark scores have been growing since October 2025 at an estimated 1.9 times the earlier rate, with a 95 per cent confidence interval of 1.5 to 2.8 times. In plain terms, Anthropic appears to have made roughly two years of historical progress in a single year.

METR’s best guess is that most of that acceleration — perhaps around 80 per cent of it — comes from Anthropic using AI to do its own research and development. That figure, the group stresses, carries substantial uncertainty. What the auditors say they can state with confidence is what did not cause the speed-up: it was not, they write, an acceleration in the scaling of training compute — the raw processing power that has driven most AI progress for a decade.

The report, published on 5 November, is the first independent measurement of AI accelerating AI research from inside a frontier laboratory. It lands at a moment of extraordinary tension in the industry: Anthropic is preparing a stock market listing that could be one of the largest ever, its chief executive has publicly called for the industry to slow down, and its great rival OpenAI has just closed the first $11 billion of a new funding round at a $1.25 trillion valuation. The audit suggests the technology is moving faster than almost anyone outside the labs can verify — and that the people building it cannot, or will not, show their workings.

What METR actually measured

METR — short for Model Evaluation and Threat Research — is a Berkeley-based nonprofit that evaluates frontier AI systems for catastrophic risk. From 12 August to 29 October 2026, about two and a half full-time-equivalent members of its staff worked inside Anthropic on what the group calls a high-access evaluation, according to METR. They reviewed internal data on pretraining and reinforcement learning, training compute in each domain, benchmark scores for released and unreleased models, internal analyses of how much AI was speeding up research, forecasts of a future model’s performance, and around 15 interviews with relevant staff. They worked on Anthropic devices and took no payment, although the company provided laptops and free tokens.

The assessment was designed to answer three questions: whether the growth rate of Anthropic’s internal capabilities has accelerated, how much of that acceleration is attributable to AI, and how quickly extreme and risky capabilities might emerge. Anthropic also asked METR to review the automated research and development section of the risk report it published in August 2026, a companion review that was published alongside the assessment.

The headline number comes from AECI, which turns results across many benchmarks into a single score. Before October 2025, Anthropic’s models gained about 15 points a year on the index. Since then, they have gained about 29 points a year, according to METR. That works out to the 1.9-fold acceleration, and METR estimates that Anthropic made roughly as much benchmark progress in the past year as it did between October 2023 and October 2025 — the span from Claude 2.1 to Claude Sonnet 4.5.

But the group is unusually candid about how much it does not know. Its subjective estimate is that the true acceleration could plausibly be as low as 1.3 times or as high as 3 times the historical rate. Robustness checks, which vary choices about which benchmarks to include, how models are dated and how saturation is handled, produce estimates ranging from 1.1 to 2.9 times. METR says it would be surprised if the true figure showed no acceleration at all, or acceleration above three times — but the width of that band is the point. This is a measurement with large error bars, made by auditors who were given remarkable access and still could not see everything.

There are other caveats. The acceleration may be limited to “well-specified tasks” — the kind that can be graded automatically or by a language model — rather than the messier work that is harder to define. The benchmark trend could reflect overfitting to the tests, or an artefact of the index itself. And the attribution of the speed-up to AI is not the result of a controlled comparison; it is an estimate built from interviews, surveys and internal analysis, not an experiment. METR notes that Anthropic’s own staff did not agree among themselves about what explained the acceleration, and that the acceleration in late 2025 — the point where the trend first bends — was likely influenced by non-AI factors, the details of which were redacted.

One of the more striking figures in the report comes from informal polls of Anthropic staff on the company’s internal messaging system. By October 2026, employees using the company’s strongest internal model reported being about 12 times as productive as they would be without AI help, according to METR — roughly double what they had reported a few months earlier. METR calls these polls very loose evidence, based on low-effort responses, and stresses that a self-reported productivity gain is not the same thing as an acceleration rate. Its own earlier work gives reason for caution: a 2025 METR study of experienced open-source developers found that AI tools actually slowed them down on the tasks tested, even though the developers felt faster.

The redactions: what the public cannot see

The most important sentence in the report is not a number. It sits in the redaction summary at the top, where METR explains that Anthropic had the right to redact any non-public information in the published document — and used it. “Anthropic redacted a majority of the evidence we found most informative for these conclusions,” METR writes, adding that the company described the redactions as aimed at removing confidential information that could accelerate other AI developers.

The cuts were not cosmetic. Anthropic refused a few of METR’s data requests outright, and cancelled a planned check of how the company splits its computing power between building capabilities and safety work. The full, unredacted report runs to about 100 pages, according to The Verge; the public version is a shadow of it. METR says plainly that the report does not achieve transparency about evidence, and that readers have to lean on the group’s judgement far more than would be ideal, given the weight of the questions involved.

That matters because the questions are not academic. Anthropic lists AI research and development acceleration as a key risk indicator in its Responsible Scaling Policy, the framework under which it is supposed to pause or slow work if certain capability thresholds are crossed. METR’s companion review found that the company’s public risk report from August “overall understates the level of acceleration that was occurring.” Whether the policy’s two-times threshold for automated research and development had been crossed by 15 July 2026 is, on the evidence available, genuinely uncertain — and the researchers involved lean in different directions, according to reporting on the assessment.

Anthropic has defended the redactions. Sam McCandlish, a co-founder of the company, wrote on X after the report appeared that disclosing the specific factors behind its progress could reveal commercially sensitive information and hand rival laboratories a roadmap for moving faster — which, he argued, would worsen the very risks METR warns about. He said Anthropic plans to publish its own analysis of the pace of AI progress, along with additional safety measures, in the coming weeks. Anthropic declined to answer The Verge’s questions about the assessment and directed the outlet to McCandlish’s post, and it did not immediately respond to a request for comment from Reuters.

The defence has a logic to it, and it also has a cost. The whole point of inviting an independent evaluator inside the laboratory was to produce evidence the public could trust. What the public has instead is a conclusion it must take largely on trust, from auditors who are themselves telling the public how much they could not show. Outside AI safety researchers quoted by The Verge questioned whether the public report provides enough evidence for independent scrutiny, and called for deeper access to frontier AI laboratories.

The boss who asked for a slowdown

The findings carry an extra charge because of who runs Anthropic. In an essay published on 12 September, chief executive Dario Amodei called on the AI industry to “slow the pace at which we improve the capabilities of AI models” so that safety measures can catch up. He defined the driving force of progress as recursive self-improvement — “AI’s growing ability to build the next generation of AI” — and warned that, “left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.”

METR’s report is, in effect, the first outside test of that claim — and it suggests the process Amodei described is already well under way inside his own company. Anthropic has said its Claude chatbot is increasingly helping the company develop newer, more advanced versions of itself, according to the Associated Press. The audit found that AI is not yet able to make highly significant, fully autonomous contributions to the most important stages of model training. But it also recorded what it called significant “signs of life” suggesting that could happen soon — and it found no convincing evidence of bottlenecks in algorithmic progress, training compute, experiment compute or data that would prevent further progress over the next 12 months.

Amodei’s essay also championed the idea of having external evaluators embedded within the leading AI laboratories, and specifically highlighted METR as an appropriate third-party evaluator. OpenAI chief executive Sam Altman said at the time that he also supported the idea of independent evaluators with employee-like access. The audit is therefore a test case for the model both men endorsed: a safety group with deep access, publishing what it can, in a company that retains the final say over what the public sees.

There is an irony that is hard to miss. The chief executive who asked the industry to slow down is running the company whose internal pace, on the only independent measurement available, has doubled. None of that makes Amodei’s warning insincere — it may make it more urgent. But it does raise an awkward question: if the head of the laboratory believes the pace is dangerous, and his own laboratory is setting that pace, who exactly is supposed to apply the brakes?

Three models nobody has seen

Buried in the reporting around the assessment is a detail that may matter as much as the headline number. Anthropic finished training three models in the past six months that it has not released, according to The Verge, completing them in July, in August to September, and in September to October — alongside an older unreleased model from 2022. METR had access to benchmark scores for these internal models, but did not get query access to the models themselves.

The existence of unreleased frontier models is normal practice: laboratories train systems, test them, and hold some back. But it is a reminder of how much of the industry’s real capability sits behind closed doors. The public judges AI progress by what is released — the chatbots, the coding tools, the features announced on stage. The audit suggests the gap between what is public and what exists inside the laboratories is widening, and widening fast.

One public glimpse of that internal capability came in the same week. A research system from a company called Physical Superintelligence, running a non-public Anthropic model, proved the Aaronson–Ambainis conjecture — an 18-year-old problem in quantum mathematics — according to AI Daily Grind. It is a small, esoteric achievement, but it is the kind of thing that was once considered the exclusive preserve of human mathematicians. It is also a concrete example of what METR means by AI accelerating research: not just writing code faster, but doing science.

The money around the finding

The audit lands in the middle of the most extraordinary financing season the technology industry has ever seen — and the money tells its own story about the pace of progress.

OpenAI closed the first $11 billion of a new funding round on Friday 6 November, at a valuation of $1.25 trillion, according to Bloomberg, in a report confirmed by Reuters. The figure is less than half of the $30 billion the company originally sought, and well below the $1.4 trillion to $1.5 trillion valuation two investors told Yahoo Finance it had initially hoped for. The round, which could still grow to about $20 billion in the coming weeks, was confirmed at the $1.25 trillion figure by OpenAI chief financial officer Sarah Friar in a television interview, although she pushed back on reports of a higher starting price. Investors in the first close include SoftBank and Abu Dhabi’s MGX, along with Coatue Management, T. Rowe Price, Sequoia Capital, Founders Fund and Andreessen Horowitz, according to The Decoder.

The scaled-back raise is the clearest sign yet that even the frontier laboratories are negotiating against a cooler market. AI stocks have sold off sharply in recent weeks after the midterm elections, as investors question the vast spending on data centres and the returns it will produce. OpenAI had spent more than a month raising after delaying its initial public offering plans, and its March round — $122 billion in committed capital at a post-money valuation of $852 billion — now looks like a product of a different market.

Anthropic, meanwhile, is pushing ahead with its own listing, potentially as soon as this month, in what could be one of the biggest initial public offerings ever. Investors have reportedly valued the company at about $1.25 trillion to $1.75 trillion for the offering, below earlier expectations of roughly $2 trillion, according to reporting on the deal. The METR audit, published days before the listing, is therefore not only a safety document. It is a disclosure that lands in the middle of a multi-trillion-dollar argument about what these companies are worth — and its redactions mean investors, like the public, are being asked to price a company whose most important asset is progress they cannot fully inspect.

Analysis: Why It Matters

The loop is the story

For years, the story of AI progress has been a story about inputs: more data, more computing power, more money. The METR assessment suggests the story is changing. The auditors say they are confident the acceleration was not caused by faster growth in training compute — the input that has dominated every previous leap. Instead, the improvement appears to come from the process itself getting better: AI systems helping researchers write code, run experiments and design the next generation of models. That is a different kind of progress, and a harder one to slow down, because it does not depend on building more data centres or buying more chips. It feeds on itself. If the loop is real, and if it compounds, then the pace of capability growth is no longer set by how much money the industry can spend — it is set by how quickly the systems can improve each other. That is precisely the scenario Amodei warned about, and the audit suggests it has already begun, even if the fully autonomous version of it has not yet arrived.

Trust is now the scarce resource

The most consequential part of this story may not be the 1.9 times figure. It may be the redactions. The AI industry’s safety argument has always rested on a promise: that the companies building the most powerful systems will let independent eyes inside, and that the public will be able to judge the risks for itself. METR was given more access than any outside group has had before — and still published a report that, by its own account, does not achieve transparency about evidence. If this is what maximum access looks like, then the public’s ability to verify safety claims is far weaker than the industry’s rhetoric suggests. That gap matters most at exactly the moments when trust is most needed: before a listing, before a deployment, before a capability threshold is crossed. A system that asks the public to trust the judgement of auditors who cannot show their evidence is a system running on reputation, not verification — and reputations, unlike benchmarks, do not come with confidence intervals.

The safety case and the business case are colliding

Anthropic’s dilemma is not unique, but it is unusually visible. The company’s public identity is built on safety: its Responsible Scaling Policy, its risk reports, its chief executive’s call to slow down. Its business reality is a stock market listing that will value it in the trillions, in a market that rewards speed. The METR report sits at the collision point. The redactions were defended on commercial grounds — disclosing the drivers of progress would help rivals — but the same disclosure would also help the public understand the risks. Every frontier laboratory faces this tension, and every one of them resolves it in favour of secrecy when it matters. The question the audit raises is whether a safety framework that depends on a company’s willingness to disclose its own risks can ever work when the company’s valuation depends on not disclosing them. Anthropic has promised its own report on the pace of progress in the coming weeks. Whether that report shows more than the redacted audit did will be a test of whether the safety case can survive contact with the listing.

The measurement problem is getting worse, not better

There is a quieter warning inside the METR report: the tools we use to measure AI progress may themselves be part of the problem. The acceleration was measured on benchmarks — tests that can be graded automatically or by a model — and METR is explicit that the finding may not extend to the messier tasks that are harder to define. Benchmarks can be gamed, overfitted and outgrown; a system that is optimised to score well on them will, by definition, get better at them. The gap between benchmark performance and real-world capability is one of the oldest problems in AI evaluation, and it grows precisely when progress is fastest. If the industry’s most important question — how fast are these systems really improving? — can only be answered with a number that carries a confidence interval from 1.5 to 2.8, a robustness range from 1.1 to 2.9, and a subjective judgement that the truth could be anywhere from 1.3 to 3, then the honest answer is that nobody outside the laboratories really knows. And the laboratories, for commercial reasons, are not telling.

The world’s regulators are watching a race they cannot see

Governments have spent the past two years building frameworks to govern frontier AI — in Washington, in Brussels, in London. Almost all of them depend on the same thing: information from the companies themselves. The Trump administration has encouraged AI companies to police themselves on safety, according to the Associated Press. The European Union’s AI Act and Britain’s AI Safety Institute both assume that laboratories will report their capabilities honestly and in time for action to be taken. The METR audit suggests that assumption is optimistic. A regulator that cannot see the evidence behind a two-fold acceleration cannot set a meaningful threshold, cannot judge whether a safety policy has been breached, and cannot act before a capability is deployed. The audit is a warning not only about the speed of progress, but about the visibility of it: the race is accelerating, and the people charged with watching it are being shown a redacted version of the track.

What to Watch Next

1. Anthropic’s own report: The company has promised its own analysis of the pace of AI progress and additional safety measures in the coming weeks. Whether it discloses more than the redacted METR audit — and whether it addresses the question of the two-times threshold in its Responsible Scaling Policy — will be the first test of the new transparency it has promised.
2. The listing: Anthropic’s initial public offering, expected as soon as this month, will be priced against a market that has just cooled on AI valuations. How investors value a company whose internal progress is accelerating while its public evidence is redacted will set the tone for every frontier laboratory that follows it to market.
3. OpenAI’s second tranche: The $11 billion first close is only part of a round that could reach $20 billion. Whether the remaining $9 billion closes at the same $1.25 trillion valuation — and whether OpenAI’s own plans for an early-2027 listing survive the sell-off — will show whether the cooling in AI finance is a pause or a turn.
4. The unreleased models: Three Anthropic models finished training between July and October 2026 without being released. When the next of them appears — and what it can do that its predecessors could not — will be the most concrete public evidence of whether the acceleration METR measured is still compounding.
5. The evaluator model itself: Amodei and Altman both endorsed embedded, independent evaluators with employee-like access. METR’s experience — unprecedented access, heavily redacted publication — will shape whether other laboratories invite the same scrutiny, and whether regulators begin to demand it instead of requesting it.

Sources

• METR — “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026”, 5 November 2026
• Reuters (via SRN News) — “Anthropic’s models improving nearly twice as fast with help from AI, study finds”, 5 November 2026
• Associated Press (via Fast Company) — “Anthropic doubles AI development amid slowdown calls”, 5 November 2026
• The Verge — “Why a new report on Anthropic automating AI development is alarming researchers”, 7 November 2026
• TechCrunch — “OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation”, 6 November 2026
• Unite.AI — “OpenAI Secures $11B in Initial Funding at $1.25T Valuation”, 6 November 2026