Blog · AI & automation

METR: Anthropic Models Improving Nearly Twice as Fast

Share: https://www.ellixorlabs.com/blog/metr-anthropic-ai-rd-acceleration-2x (copy URL for iMessage or social)

Published (display in IST)

Updated (IST)

METR cover graphic for its Anthropic AI R&D acceleration assessment

METR's 5 November 2026 review finds Anthropic likely improving well-specified task capabilities at about 2x the prior rate, mostly from AI-assisted R&D rather than faster training compute.

Image: METR cover graphic for its Independent Assessment of AI R&D Acceleration at Anthropic in October 2026. View source.

Independent evaluators at METR say Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025, according to a public report published on 5 November 2026.

The eleven-week review, from 12 August to 29 October 2026, used internal benchmarks, staff interviews and other private Anthropic data. METR estimates Anthropic's AECI benchmark trend has grown at about 1.9 times the pre-October 2025 rate, with a 95 percent confidence interval of 1.5 to 2.8 times. Channel NewsAsia, citing the same study, reports the finding as nearly twice as fast, and The Verge independently summarised the headline conclusion.

Direct answer: METR concludes Anthropic's recent capability growth on well-specified tasks is likely about 2x historical pace, mostly from AI-assisted R&D rather than faster training-compute scaling, with high uncertainty and heavy redactions.

What METR found

METR, also known as Model Evaluation and Threat Research, framed three questions: whether Anthropic's internal capability growth accelerated, how much of any acceleration comes from AI, and whether clear bottlenecks would slow further gains. Most METR staff on the project now believe the answer to the first question is roughly a doubling versus the March 2024 to October 2025 baseline, though they say they would not be surprised if the true figure sat nearer 1.3x or 3x.

On attribution, METR says it is confident the speed-up was not caused by faster growth in training compute. Its subjective guess is that a majority of the October 2026 acceleration, perhaps around 80 percent, is due to AI use, with substantial uncertainty. Interviews and surveys were consistent with AI usage driving a doubling of progress pace. Informal Slack polls of self-reported uplift reached about 12x for the most capable internal model in October 2026, but METR calls that evidence very loose and hard to interpret.

AI is not yet making highly significant, fully autonomous contributions to the most important stages of model training, METR writes. Researchers did describe "signs of life" that such contributions could matter soon. The team also did not identify convincing bottlenecks in algorithms, training compute, experiment compute or data that would clearly bite within the next twelve months, while warning that its view of those bottlenecks changed several times during the review.

Caveats, redactions and company context

Anthropic held unilateral redaction rights over non-public material. METR says Anthropic redacted a majority of the evidence the evaluators found most informative for their risk conclusions, describing the cuts as aimed at confidential information that could accelerate other AI developers. That leaves the public more dependent on METR's judgment than the authors would prefer.

Robustness checks on AECI still cluster near 2x under many snapshot and benchmark choices, though link-function assumptions swing estimates from about 1.1x to 2.9x. METR also lists ways it could be over- or under-estimating real-world impact, including the focus on well-specified, easily graded tasks.

The Verge notes that METR engaged Anthropic from mid-August and submitted the final public report on 3 November 2026, after CEO Dario Amodei's mid-September essay urging the industry to pace frontier progress. Anthropic co-founder Sam McCandlish said the company plans to share its own analysis of the speed of AI progress within Anthropic. Channel NewsAsia adds that Anthropic did not immediately respond to a request for comment and is preparing for a possible public listing.

What it means for builders

For product and engineering teams, the practical signal is not a single multiplier. It is that a frontier lab's measured progress on clean tasks appears to have steepened while staff lean harder on AI for coding, experiments and analysis.

If you ship agents or internal copilots, expect competitor capabilities and open tooling to move faster than last year's planning cycles assumed. Budget more frequent eval refreshes, tighter tool-use monitoring and clearer human review gates when models get long-horizon autonomy. Those lessons already show up in nearby safety and security stories, including OpenAI's GPT-6 Astra findings and the GPT-6 Cyber release.

Teams that need help designing evaluation harnesses, agent guardrails or AI product surfaces can see Ellixor's AI and automation services.

What to watch next

Watch Anthropic's promised internal progress report, any update to its Responsible Scaling Policy indicators, and whether later METR or third-party reviews confirm sustained acceleration once more evidence is public. Also watch how IPO timing and compute deals interact with claims that AI itself is now a major input to AI R&D.

For related Anthropic product context this week, see Claude Personal Finance, and for nearby model launches compare Thinking Machines' Inkling 2 Small. For a build brief or eval plan, contact Ellixor.

Frequently asked questions

What did METR conclude about Anthropic's AI progress rate?

METR's 5 November 2026 public report says Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025, with high uncertainty. AECI benchmark growth is estimated at about 1.9x the earlier trend.

Is Anthropic's faster progress caused by more training compute?

METR says it is confident the observed acceleration was not due to faster scaling of training compute. Its subjective estimate is that a majority of the October 2026 acceleration is attributable to extensive use of AI for research and development, though uncertainty remains high.

Can Anthropic's AI systems fully automate model training yet?

No. METR writes that AI is not yet making highly significant, fully autonomous contributions to the most important stages of Anthropic's model training process. Researchers described signs of life that such contributions could play a larger role soon.

FAQ

What did METR conclude about Anthropic's AI progress rate?

METR's 5 November 2026 public report says Anthropic is likely improving model capabilities on well-specified tasks at around twice the rate seen between March 2024 and October 2025, with high uncertainty. AECI benchmark growth is estimated at about 1.9x the earlier trend.

Is Anthropic's faster progress caused by more training compute?

METR says it is confident the observed acceleration was not due to faster scaling of training compute. Its subjective estimate is that a majority of the October 2026 acceleration is attributable to extensive use of AI for research and development, though uncertainty remains high.

Can Anthropic's AI systems fully automate model training yet?

No. METR writes that AI is not yet making highly significant, fully autonomous contributions to the most important stages of Anthropic's model training process. Researchers described signs of life that such contributions could play a larger role soon.

Talk to Ellixor Labs about this topic.

Source: METR