Anthropic AI research may be advancing at roughly twice its earlier pace, according to an outside assessment published on 5 November 2026. The non-profit evaluation group METR was given unusually detailed access to the company’s research process and concluded that progress on certain well-specified model tasks appears substantially faster than before. In its published assessment, METR also said a majority of that acceleration is likely related to Anthropic using AI to help develop more capable AI.

That sounds like the beginning of a machine improving itself without human intervention. The report does not say that. METR describes extensive AI assistance to researchers, while finding that fully autonomous contributions to the most important parts of model development are not yet a large driver. The distinction is crucial: a faster human-led development process is serious news, but it is not proof that an independent AI research system has taken control.

What the Anthropic AI research assessment actually measured

METR examined internal information between 12 August and 29 October 2026, including model evaluations, the research process and interviews with staff. Its central estimate is that capability progress on certain clearly defined tasks is happening at around twice the historical rate observed in an earlier period. The group emphasises high uncertainty: depending on interpretation, it would not be surprised by a materially lower or higher rate.

A separate measurement using Anthropic’s internal benchmark series suggested about 1.9 times the previous pace. Benchmarks are structured tests, not a direct measure of how a system performs in every workplace or household. A model can improve quickly on defined tasks without becoming equivalently capable at an unfamiliar real-world job. METR also explored alternative ways of analysing the data and acknowledged that methodological choices affect the estimate.

Its most striking inference concerns cause. The evaluator said faster growth in training computing resources was not the explanation for the observed acceleration. It judged that AI tools doing useful work for researchers were probably contributing a large share of the increase. Staff interviews and internal indicators supported that possibility, although interviews and self-reports do not offer the certainty of a controlled experiment.

The evidence the public has not been allowed to inspect

The report contains a serious qualification. Anthropic retained rights to remove non-public material, and METR says the company redacted much of the evidence most informative to its conclusions. The evaluators were permitted to disclose the fact of those redactions, but the public cannot independently weigh important parts of the case. That limits what readers should conclude from a striking headline about the speed of progress.

METR’s task was also narrower than a comprehensive audit of the company. The group explains that a planned component examining the allocation of computing resources was cancelled partway through the engagement. It should therefore not be presented as proof that Anthropic devotes any particular share of its infrastructure to safety or capabilities. A detailed external assessment can still leave significant unanswered questions.

LiveAIWire has previously looked at the difficulty of interpreting benchmarks about AI improving AI. The present report illustrates that problem in unusually concrete terms. It contains evidence of a real change in research productivity, but it cannot neatly translate that change into a date when a system might handle an entirely new category of work.

Why the pace matters even without runaway AI

If researchers can use more capable systems to design experiments, inspect code and accelerate their own development work, progress can compound. The safety question is whether assessment, testing and governance can accelerate as well. A company making stronger systems faster may need a stronger process for deciding when capabilities become too risky for ordinary release or when further safeguards are required.

Anthropic has described those questions in its Responsible Scaling Policy. Such policies are company commitments rather than substitutes for independent law or enforcement. They may help define what a company intends to measure, but their credibility depends on transparency, independent scrutiny and whether their safeguards work under pressure.

There is a financial context too. LiveAIWire has reported on Anthropic’s proposed stock-market flotation. The timing makes credible analysis of capability progress particularly important: researchers, potential investors, policymakers and customers all need something firmer than company optimism or public fear. METR’s work is valuable precisely because it lays out both findings and limits.

The pace of progress also depends on what is being measured. A benchmark that asks a model to complete a clearly specified research task can reveal important changes over time, but it does not capture every stage of producing dependable scientific work. Researchers still choose problems, decide which evidence is convincing and determine whether an experiment really answers the question. A useful AI assistant can speed up parts of that process without taking responsibility for the whole.

That matters for interpreting a possible feedback loop. Faster development can produce a more capable model, which might then help researchers with the next round of development. METR’s assessment raises that possibility as a serious subject for monitoring. It does not establish an inevitable acceleration curve, and the report itself warns that future bottlenecks may be difficult to foresee. Treating one estimate as a timetable would go beyond the available evidence.

The conclusion is neither that a runaway intelligence has arrived nor that the report can be dismissed as another benchmark story. The credible finding is a possible marked acceleration in a major AI laboratory’s development process. The equally credible caution is that the magnitude, mechanism and future direction are uncertain, with key evidence unavailable for public review. Both facts belong in the same headline conversation.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.