AI research is speeding up, and it is no longer just a feeling. According to an independent report published on November 5, 2026 by the nonprofit METR (Model Evaluation & Threat Research), the pace of progress in Anthropic's models has roughly doubled since late 2025, an acceleration that METR attributes to nearly 80% to AI itself, which now takes part in its own research. For a small business owner, this figure is not just a technical curiosity: it signals that AI tools will be replaced faster than expected, bringing both opportunities and risks.
In brief
- METR, an independent AI model evaluation nonprofit, estimates that Anthropic's capability growth rate has roughly doubled (range of 1.3x to 3x) between October 2025 and October 2026, compared with the March 2024 to October 2025 period.
- An internal metric, the AECI (Anthropic Epoch Capabilities Index), has grown 1.9 times faster since October 2025, equivalent to about two years of historical progress compressed into a single year.
- METR attributes roughly 80% of this acceleration to Anthropic's own researchers using AI tools to speed up their research work, rather than to an increase in available computing power.
- The assessment ran from August 12 to October 29, 2026, with near-employee-level access for METR to Anthropic's teams and tools.
- METR notes that Anthropic redacted a significant share of the evidence before publication, and that its public report probably understates the true scale of the acceleration.
Who is METR, and what exactly does this report measure?
METR is an independent nonprofit based in Berkeley, specializing in evaluating the capabilities and risks of the most advanced AI models. LUWAI already introduced it in October 2026, when Anthropic granted it near-employee-level access to its offices (see our article on independent AI evaluators). This new report, titled "Independent Assessment of AI R&D Acceleration at Anthropic," is the first concrete output of that close collaboration.
METR's objective was specific: measure whether AI genuinely helps Anthropic's researchers build better models, and how fast. To answer that, a METR team observed Anthropic's research teams for eleven weeks, with access to internal data, working tools, and performance metrics that are usually not public.
How AI ends up accelerating its own research
The mechanism METR describes resembles a factory using its own robots to build better robots. In practice, Anthropic's researchers rely on AI agents to write experimental code, test hypotheses, and analyze results, tasks that used to take weeks of human work.
Delegate repetitive tasks
Explore more hypotheses
Get a more capable model
Restart the loop
METR notes that it found no clear sign of this loop slowing down at this stage: no identified data shortage or compute limit that would hold back this acceleration in the coming months.
The numbers behind the scale shift
To put the scale of the phenomenon in context, METR compares two periods: the baseline phase (March 2024 to October 2025) and the recent phase (since October 2025).
| Metric | Before (March 2024 - Oct. 2025) | After (since Oct. 2025) |
|---|---|---|
| Capability growth rate | Baseline (1x) | 1.3x to 3x (median estimate 2x) |
| AECI index growth | Baseline | 1.9x (confidence interval 1.5 to 2.8) |
| Share attributed to AI itself | Not measured | ~80% |
| Share attributed to more compute | Main historical driver | Ruled out by METR as a cause of the acceleration |
In other words, according to METR, roughly two years of normal progress have been compressed into a single year. It is not that Anthropic invested massively in more chips: it is that research teams now work alongside AI collaborators that speed up their own work.
The gray areas METR itself acknowledges
In the interest of honesty, it is also worth noting what this report does not guarantee. METR says so explicitly: Anthropic redacted a significant share of the evidence before publication, for commercial confidentiality reasons. METR therefore believes its own public report probably understates the true scale of the acceleration, since it could not document everything.
Key takeaway
According to METR, the productivity gains reported by Anthropic's own teams remain "weak evidence": they are internal statements, not fully independent measurements. What's more, the most critical research steps are not yet fully automated: humans remain necessary to check and correct the AI's work.
A concrete example illustrates this caution: in a separate biology research project run by Anthropic with Claude, an AI-generated protocol included an unsuitable additive in a bacterial growth medium, a mistake that cost two weeks of work before a human researcher caught and fixed it. Acceleration does not remove the need for supervision, it changes its nature.
What this acceleration means for a small business
A small business has neither the research teams nor the scale of Anthropic. But METR's finding has a direct consequence for any company relying on AI tools: the pace at which models get replaced is likely to speed up, which changes how to invest in these tools.
Frozen AI strategy
Agile AI strategy
In practice, this means avoiding building critical processes around the current limits of a single model, since those limits can move fast. It also means keeping an eye on independent audits (like METR's) rather than relying solely on vendors' marketing announcements, and continuing to check the work AI produces, even as models improve.
FAQ
What is the METR report on AI research acceleration at Anthropic?
It is an independent assessment, published on November 5, 2026, that measures how much AI now helps Anthropic's researchers build better models. METR carried out this assessment with internal access to Anthropic's teams and tools, from August 12 to October 29, 2026.
Will the pace of AI progress keep doubling every year?
No one can say for certain. METR only states that it found no clear sign of this acceleration slowing down at this stage, without guaranteeing it will continue at the same pace in the following years.
Can you trust a report redacted by the lab it is evaluating?
METR itself acknowledges the limits of the exercise: some evidence was removed by Anthropic for commercial reasons, which leads METR to estimate that its report probably understates the real acceleration. It is an independent assessment, but a partially constrained one, rather than a fully open audit.
What should a small business change in its AI strategy given this acceleration?
It should mainly avoid locking its processes around the current limits of a single model, favor interoperable tools, and regularly review its AI tool choices instead of keeping an old setup without checking it.
Conclusion
The 2026 METR report confirms that AI research has entered a feedback loop where AI contributes to its own improvement, with a pace of progress roughly twice as fast as before late 2025. For a small business, the right response is not panic but agility: keep AI tool choices reversible, follow independent audits rather than marketing announcements alone, and keep supervising the work AI produces. To go further on independent evaluators, read our article on METR and Anthropic, and for budgeting your AI tools, our guide on the real cost of AI agents.
