Anthropic AI development is moving nearly twice as fast as it did a year ago — and the biggest reason, according to an independent study released Thursday, is the company’s own artificial intelligence. The Berkeley-based nonprofit Model Evaluation and Threat Research, known as METR, found that the pace of Anthropic AI capability gains has roughly doubled even while CEO Dario Amodei publicly urges the industry to slow down.
The finding lands weeks before Anthropic plans what could be the biggest initial public offering in stock-market history. It also carries unusual credibility: METR reviewed internal data spanning about eleven weeks, interviewed staff, and says it confidently ruled out the most convenient explanation — simply spending more on computing power.
Table of contents
- Anthropic AI Acceleration: What METR Actually Measured
- Eleven Weeks Inside: How the Anthropic AI Study Was Conducted
- Recursive Self-Improvement: The Engine Behind Faster Anthropic AI
- The Slowdown Paradox: Amodei Preaches Brakes, Data Shows Accelerator
- An IPO at Full Speed: Markets Meet the METR Numbers
- Redactions and Caveats: What the Anthropic AI Report Could Not Say
- Independent Oversight: A New Bargain Between AI Labs and Watchdogs
- What Anthropic AI Acceleration Means for Everyone Else
- References
Anthropic AI Acceleration: What METR Actually Measured
METR’s central finding is blunt: measured on well-specified software tasks, Anthropic AI progress this year is arriving at roughly double the speed of twelve months ago. The research organization, which has evaluated frontier models since 2023, gained what it described as unprecedented access to Anthropic’s internal development data and interviewed the engineers working inside the process.
Just as important is what METR ruled out. The most comfortable explanation for the jump — that Anthropic simply threw more computing power at training — was dismissed with confidence. Compute growth could not account for the curve. Something in the development process itself changed, and METR’s leading explanation is that Anthropic’s own models are now doing more of the work of building Anthropic AI.
Translated into engineering terms, the Anthropic AI doubling means a research team that previously needed a year to move a benchmark metric by a given amount now achieves the same gain in roughly six months. Hiring has not doubled on the same trajectory, and internal timelines were not compressed by executive edict. The extra speed comes from tools — Claude drafting training code, reviewing evaluation output, and proposing the next experiment — stacked across thousands of concurrent projects.
That compounding is quiet and unglamorous. No single milestone announces it; the curve bends because hundreds of small tasks now finish in hours rather than days. METR’s reading is that Anthropic AI has crossed the point where its own models materially reshape the research workflow — the first frontier lab for which an outside auditor could make that claim with internal data in hand.
| Item | Details |
|---|---|
| Evaluator | Model Evaluation and Threat Research (METR), a Berkeley-based nonprofit |
| Review period | Mid-August to late October 2026 — roughly eleven weeks of internal access |
| Headline finding | Anthropic AI development moving nearly twice as fast as a year ago on well-specified tasks |
| Leading explanation | AI-assisted research and development — Claude tools building Claude |
| Compute growth | Confidently ruled out as the primary driver by METR reviewers |
| Context | Published weeks before Anthropic’s planned IPO, potentially the largest ever |
To see the scale of the shift, compare the pace METR recorded across three eras. From October 2023 to October 2025, capability gains followed the steady cadence most observers had grown used to. The current year looks like something else entirely — the equivalent progress compressed into half the time.
Eleven Weeks Inside: How the Anthropic AI Study Was Conducted
METR framed its assessment around two questions: whether the growth rate of Anthropic AI capabilities had genuinely accelerated, and how much of that acceleration was attributable to AI itself. To answer them, the researchers were granted what the group called “unprecedented access to internal data” on the company’s research and development process, and they interviewed the staff doing the work.
That access window — roughly eleven weeks from mid-August through late October 2026 — produced one of two reports the nonprofit published. The timing was not incidental: Anthropic invited METR in as the company prepared its Wall Street debut, amid swirling debate over whether advanced AI models could one day elude human control with catastrophic consequences.
METR is not an outsider throwing stones. The Berkeley-based nonprofit has conducted capability evaluations for Anthropic, OpenAI, and other leading laboratories for years. It is funded through donations and takes no compensation for its work; the AI companies it partners with provide access and tokens, which METR uses for evaluations, research, and engineering.
The framing matters for everything that follows. Because METR’s team has spent years benchmarking the same labs on comparable tasks, its read on Anthropic AI’s trajectory rests on a continuous measurement series rather than a one-off snapshot — which is precisely why a doubling finding carries weight even when the report around it stays cautious.
Recursive Self-Improvement: The Engine Behind Faster Anthropic AI
Anthropic has said openly that its Claude chatbot increasingly helps the company develop newer, more advanced versions of itself. METR’s researchers landed on the same mechanism: after ruling out compute growth, they guessed the improvement came from using AI itself to improve AI. In other words, the leading explanation for the Anthropic AI acceleration is Anthropic AI.
This is recursive self-improvement in a mundane, industrial form — not a sci-fi takeover, but thousands of engineers and product teams accelerating the research pipeline with models that write code, draft evaluations, and triage experiments. We covered the risk dimension of this loop in our piece on Anthropic’s warning about self-building AI, and METR’s data suggests the loop is already spinning faster than its own CEO’s public framing implies.
The phrase recursive self-improvement conjures a singular artificial intelligence redesigning itself overnight. The Anthropic AI reality is administrative: model-assisted code review, automated experiment triage, evaluation scripts written by the very systems being evaluated. METR’s point is that the mundane version is precisely what accelerates a capability curve — thousands of ordinary hours removed from the pipeline of model creation, every week.
It also changes who counts as a bottleneck. When senior researchers spend their days directing capable assistants, experience stops being the constraint on how fast new experiments launch. That is why Anthropic AI’s velocity could keep climbing even on a flat headcount — and why METR expects the acceleration to persist rather than spike and fade.
METR was careful not to overclaim. “We understand that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process,” the report states. Yet Anthropic’s own researchers described significant “signs of life” suggesting fully autonomous contributions could arrive soon.
The concern is compounding. Using AI to improve AI could accelerate the acceleration itself, and METR found no bottlenecks — no compute restrictions, no data shortages — currently holding Anthropic AI progress back. An engine with no governor is what independent measurement is for.
The Slowdown Paradox: Amodei Preaches Brakes, Data Shows Accelerator
The timing of METR’s access produced one of the stranger juxtapositions in recent AI history. During the very eleven-week window in which reviewers watched Anthropic AI development double in speed, CEO Dario Amodei published a September 12 essay urging the industry to “slow the pace at which we improve the capabilities of AI models” so safety measures can catch up.
The essay, written in response to mounting concerns about runaway AI, has already been the subject of our own analysis — see our breakdown of Amodei’s call to pace AI development. METR’s numbers now let readers test that call against observed reality: no one asked the company to speed up, and its internal velocity doubled anyway, driven by the very capability gains Amodei warned about.
The paradox is not hypocrisy — it is physics. When every part of the research pipeline gets faster because models assist their own creators, slowing down requires a coordinated act of restraint no single lab can perform alone. Amodei himself named the underlying concept in his essay: recursive self-improvement, or “AI’s growing ability to build the next generation of AI.”
“Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all,” he wrote. METR’s report is, in effect, a measurement of exactly how far the capability frontier has outrun oversight since those words were published.
An IPO at Full Speed: Markets Meet the METR Numbers
The San Francisco-based company plans to go public later this month in what could be the largest initial public offering in stock-market history. The METR analysis — evidence that Anthropic AI development is compounding — now sits in front of investors, regulators, and the public in the same news cycle, and Anthropic declined to comment on the findings when approached by the Associated Press.
Investors reading the report see a research machine whose own tools are making it faster. Skeptics see a company racing toward public markets while its CEO publicly asks an industry he cannot unilaterally brake for patience. Both readings are on the table at open.
| Date | Event |
|---|---|
| October 2023 | METR’s benchmark evaluation era begins — the baseline period for measuring capability gains |
| Mid-August 2026 | METR’s eleven-week internal review of Anthropic AI research begins |
| September 12, 2026 | Amodei publishes essay urging the industry to slow capability improvements |
| November 5, 2026 | METR releases findings: Anthropic AI development roughly twice as fast as a year ago |
| November 2026 | Anthropic plans IPO that could be the biggest ever |
The ecosystem around the numbers is growing too: third-party verification startups are positioning themselves to audit AI claims for a public now expected to scrutinize Anthropic like any listed company — our profile of FlashProve’s cryptographic model proofs for Anthropic shows what that market looks like in practice.
Listing forces a different kind of disclosure: quarterly numbers, risk factors, and lawyers who read METR’s report before the roadshow slides are drafted. The Anthropic AI doubling finding will sit permanently adjacent to the prospectus, if only as a named risk. No frontier lab has taken this particular speedometer public before.
Redactions and Caveats: What the Anthropic AI Report Could Not Say
METR itself flagged the limits of its headline. The report was peppered with caveats: the nonprofit said it could be overstating its findings, because the measured acceleration might apply only to “well-specified tasks” that can be graded automatically or by a language model, rather than to “messy” tasks that resist clean definition. It is also possible, METR allowed, that Anthropic understated its earlier progress — making the latest jump look bigger than it was.
Anthropic’s redactions to the report further constrained what the researchers could communicate. The redactions limited how precisely METR could express its estimates of acceleration, and the team carried internal disagreements and uncertainty through the review. Even so, the group concluded the company’s rate of improvement had doubled.
| Factor | Well-specified tasks | Messy tasks |
|---|---|---|
| How they are graded | Automatically or by a language model | Human judgment, hard to define success |
| Measurement confidence | High — continuous METR benchmark series | Low — too ambiguous to track the same way |
| Acceleration evidence | Strong — roughly 2x year over year | Unknown — acceleration may not extend here |
| Risk of overstatement | Possible if earlier progress was understated | The doubling might not generalize at all |
Honest numbers with visible limits beat confident numbers without them. The doubling claim is real but bounded: it describes the part of Anthropic AI work that machines can grade, and METR explicitly does not promise the same speedup where humans, not graders, define success.
Independent Oversight: A New Bargain Between AI Labs and Watchdogs
The findings carry weight for a structural reason: details about advanced AI’s technology and risks have come primarily from the AI companies themselves, encouraged by President Donald Trump’s administration to police themselves on matters of safety. An independent group publishing hard numbers about a lab’s internal velocity is a deviation from that pattern — and an experiment in what replaces self-reporting.
According to the Associated Press’s account of the METR analysis, the nonprofit is funded through donations and does not accept compensation for its evaluations; AI companies it partners with provide access and tokens, which METR spends on its own evaluations, research, and engineering. That model keeps the watchdog independent of grant cycles while making it dependent on lab goodwill for access — a tension worth watching as these arrangements multiply.
The bargain may be deepening. In his September essay, Amodei championed the idea of external evaluators embedded within leading AI labs and specifically named METR as an appropriate third party. OpenAI CEO Sam Altman said at the time that he too supported independent evaluators with employee-like access.
Two of the strongest labs in the world endorsing the same oversight mechanism is either the beginning of a norm or a sophisticated piece of positioning — and METR’s Anthropic numbers give both interpretations something to feed on.
What Anthropic AI Acceleration Means for Everyone Else
If METR is right, the Anthropic AI doubling is not a story about one company’s discipline or lack of it. It is evidence that when models become useful at building models, capability growth accelerates on its own schedule — responsive to tooling and talent, largely indifferent to public letters. A single firm’s call for caution does little against an industrial process whose main input is now the previous generation of its own product.
The compounding risk METR named is the one to keep in view: AI improving AI could accelerate the acceleration, and no bottleneck — compute, data, funding — currently binds Anthropic’s progress. The first constraints, if history is a guide, will come from outside the lab: from markets, from watchdogs with data, or from an incident.
First, the IPO itself: whether public-market disclosure produces a fuller picture of internal capability trends than any voluntary METR review can. Second, whether the embedded-evaluator idea Amodei and Altman endorsed becomes a standing arrangement with published findings, or fades once fundraising is done.
Third, whether METR’s method spreads — other labs inviting the same eleven-week look, producing a continuous, independent speedometer for the whole frontier. If Anthropic AI’s doubling is the first reading on that instrument, it will not be the last.
Other laboratories will now be asked — by investors, by regulators, by their own staff — for the kind of independent velocity review METR performed on Anthropic AI. Whether that demand produces a standard or theater depends on whether the reports keep their caveats and whether the access survives the next fundraising cycle.
For everyone outside the frontier labs, the practical lesson is simple: assume the doubling is real for the tasks that can be graded, stay skeptical about the rest, and expect the next measurement — at another lab, within months. The Anthropic AI case has set the template; the queue behind it is long. This machine does not idle.
References
AP: Anthropic is developing AI around twice as quickly, monitoring group says
Reuters: Anthropic’s models improving nearly twice as fast with help from AI, study finds
METR: R&D Section Anthropic Risk Report — August 2026 Review
METR: Model Evaluation and Threat Research
Mercury News: Anthropic’s AI development accelerating, METR finds
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
