Roughly Two Years of Progress in One: METR Finds Anthropic Standing at Its Own 2x Threshold

Independent evaluator METR published its assessment on November 5, 2026 of how quickly Anthropic is using artificial intelligence to accelerate its own research.

File photo: File photo of Anthropic's San Francisco headquarters during a visit by UK Secretary of State Peter Kyle in October 2024.
File photo: Anthropic's San Francisco headquarters during a visit by UK Secretary of State Peter Kyle in October 2024. File photo: Department for Science, Innovation and Technology, CC BY 2.0 · source

Roughly Two Years of Progress in One: METR Finds Anthropic Standing at Its Own 2x Threshold

Independent evaluator METR published its assessment on November 5, 2026 of how quickly Anthropic is using artificial intelligence to accelerate its own research.

Independent evaluator METR published its assessment on November 5, 2026 of how quickly Anthropic is using artificial intelligence to accelerate its own research. Its conclusion: the company's models are likely improving at roughly twice the historical pace of progress — almost exactly at the threshold Anthropic itself has defined as "dramatically accelerated progress" in its own safety policy. At the same time, the company has changed the language of its IPO filings, and the week brings an IPO reportedly valuing the company at up to $1.4 trillion. And most of the evidence underlying the conclusion is redacted.

This is not a story about an alarm line being definitively crossed. It is a story about the boundary between "normal scaling" and Anthropic's own safety regime now being adjudicated under considerable uncertainty — by a company in the process of selling itself to the public market, with an independent assessment the public must partly take on faith.

What METR actually measured

From August 12 to October 29, 2026, METR staff conducted a high-access evaluation inside Anthropic, measuring the extent to which AI accelerates AI research. According to METR, the group received "unprecedented access to internal data on their AI research process," including pretraining and reinforcement learning metrics, training compute per domain, benchmark scores for internal models, analyses of AI-generated speedups, forecasts of a future model's metrics, and interviews with relevant personnel (METR).

The headline number is this: as of October 2026, Anthropic is likely improving the capabilities of its models on well-specified tasks at roughly twice the rate seen from March 2024 to October 2025. METR states the uncertainty itself: "We would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate."

The most concrete measure is benchmark growth. Measured through AECI — Anthropic's version of the Epoch Capabilities Index — benchmark scores for Anthropic's models have grown since October 2025 at an estimated 1.9x the pre-October 2025 rate, with a 95 percent confidence interval of 1.5 to 2.8x. That means, METR writes, that Anthropic has achieved "roughly two years of historical progress in the past year." METR also notes that Anthropic's private risk report implies growth of 1.97x, while other "reasonable estimates" range from 1.3 to 2.8 depending on which metric is used (MSN/IBD).

The mechanism: AI as researcher, not just computation

Perhaps the most important finding is not the pace but its source. According to METR, the majority of the acceleration is now likely due to Anthropic's extensive use of AI to carry out the research work itself — not to scaling of compute. The group reports that roughly 80 percent of October's growth came from AI systems assisting AI research, rather than expanded compute capacity, while claiming at the same time that Anthropic is understating part of the acceleration (MSN/IBD).

METR adds a significant caveat: there is "significant uncertainty," and other causes likely played a large role in the acceleration as late as the end of 2025. In other words, the dividing line between "AI is driving the research" and "more GPUs and better methods" is not drawn with surgical precision — and it is precisely this line that determines whether the policy threshold has been reached.

What the threshold triggers

Anthropic's Responsible Scaling Policy defines a "2x Acceleration Threshold": if progress attributable to AI becomes "dramatically accelerated," a series of safety commitments come into force. According to Gizmodo's rendering of the policy document, the company must then explore "ways to achieve unprecedented levels of security," gain full visibility into what its models are doing — a condition the policy calls "eyes on everything" — and learn to red-team itself so thoroughly that bug bounty programs become redundant (Gizmodo).

This is not an external regulation but a promise the company has made to itself — and which is now being put to the test by an independent evaluation group.

The S-1 timeline: from "may soon" to silence

In the version of Anthropic's S-1 dated November 2, the company stated that it "may imminently meet" the criteria its Responsible Scaling Policy sets for "dramatically accelerated progress attributable to AI." In the amended version of November 10, this language is removed, and the remaining language suggests, according to Gizmodo, that the policy's relevant provisions have taken effect.

Precision matters here: this is Gizmodo's reading of the SEC filings, and it is partly inferential — the reader is left to draw the conclusion about which provision of the policy applies. None of the sources report Anthropic explicitly saying that the threshold has been crossed or that the obligations have been triggered. What the company has actually said publicly is, in sequence: first that the threshold may be reached "soon," then a revision in which the language is gone. Between those two points lies METR's assessment that the company is at "around" 2x.

The timing collision: an IPO in the same week

The company is set to price its IPO on Wednesday and begin trading on Thursday, under the ticker ANT on the Nasdaq Global Select Market. The valuation — reportedly up to $1.4 trillion — is reported, not confirmed by Anthropic (MSN/IBD).

The prospectus contains a striking warning: "these risks will increase as AI systems substantially automate AI research and development" (The Atlantic). In other words: the company is selling shares with a business plan that depends on precisely the development now approaching its own safety threshold — and the warning language in the document points in the same direction as METR's findings.

The surrounding context is also worth noting. On September 12, CEO Dario Amodei published an essay urging the industry to "slow down the pace of improving AI model capabilities" so that safety work can keep up (Dealroom). Two months later, an independent evaluation of his own company shows it driving roughly twice the historical rate of capability growth.

The dispute: the August evaluation versus METR

The most direct conflict concerns a date: July 15. In its own safety evaluation in August, Anthropic concluded that the risk of the 2x acceleration threshold having been crossed by that date was low. METR disagrees:

"We consider the argument in the public report that the 2x acceleration threshold was not crossed to be overstated, and consider it plausible that this threshold was crossed by July 15," METR writes (MSN/IBD).

This is more than an academic question. If the threshold was crossed in mid-July, the company has operated for months without the obligations the policy says should take effect — unless they have in practice already been followed without being explicitly communicated. The estimates underlying the question range, per METR, from 1.3x to 2.8x depending on metric, and Anthropic's own private risk report implies 1.97x — a figure above the threshold, while the public August report argued it had not been reached.

The transparency problem

Here lies the story's most uncomfortable structure: METR can share its conclusions but not its evidence. "While we could freely share our high-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions," the group writes (METR).

That means the public — investors, policymakers, researchers — must take METR's judgment on faith where the most decisive data are concerned. It also means METR's own uncertainty is harder to assess independently. The group is open about its limitations: it saw no signs of significant bottlenecks that would hinder further capability progress over the next 12 months — in algorithmic progress, training compute, experiment compute, or data — but adds that "our understanding of the status of various bottlenecks changed multiple times as we received more detailed information, so we are not confident in our current assessment."

What remains unresolved

It is worth summarizing what is actually known, and what is not.

Known: METR estimates that Anthropic's capability growth is around 2x the historical rate, with a primary measured figure of 1.9x and wide uncertainty. The majority of the acceleration is likely due to AI-driven research. Anthropic and METR disagree on whether the threshold was crossed by July 15. Anthropic has redacted most of the most informative evidence. The language of the IPO filings changed between November 2 and November 10.

Unresolved: whether the threshold has actually been crossed. Whether the Responsible Scaling Policy obligations have formally taken effect — none of the sources report an explicit statement from Anthropic on this. Whether the increased security requirements are actually being met in practice — "unprecedented levels of security," "eyes on everything," extensive self red-teaming — cannot be verified with publicly available information. And the reported valuation of up to $1.4 trillion is not confirmed.

The nearest point at which this may be clarified is if Anthropic itself, after the IPO, states clearly that the threshold has been reached — and describes which concrete measures follow. Until then, the public is left in a paradoxical position: a company about to be valued at potentially historic sums, on the basis of a capability trajectory that its own independent evaluation group places at its own risk boundary — and an evidence base the public largely cannot see.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.