METR estimates 1.9x benchmark-progress acceleration, with a 1.5–2.8 interval.
The report says current acceleration is mainly AI-assisted, not a demonstrated autonomous R&D loop.
What happened?
METR published an independent assessment on November 5 after reviewing Anthropic’s research process from August 12 through October 29, 2026. The assessment asks a narrow but consequential question: whether AI is accelerating the work used to develop better AI systems. METR says its access included internal benchmark scores, training and reinforcement-learning information, process metrics, interviews, surveys and forecasts. It also says Anthropic exercised redaction rights over a majority of the evidence METR found most informative, so the public cannot audit the most sensitive parts of the case.
The headline estimate is not that Anthropic’s models are “twice as intelligent.” METR says benchmark scores in its version of the Epoch Capabilities Index grew at an estimated 1.9 times the pre-October 2025 rate, with a 95% interval of 1.5 to 2.8. METR’s broader subjective conclusion is that capability progress on well-specified tasks may be around twice the earlier rate, but it would not be surprised by a true value closer to 1.3 or 3. The report attributes the estimate to growth rates, not to a single model launch or a direct measure of economic output.
The report identifies internal Models 2, 3 and 4 in its timeline, but the public page does not provide a complete, independently checkable model-by-model table against Opus 5.5. METR also says it did not receive sample access to non-public models and that a planned compute-allocation verification was canceled after Anthropic changed scope. Its evidence is therefore a high-access assessment with important limits, not a public benchmark release. This is a rate estimate, not a promise of future growth. Most importantly, METR reports that current R&D gains are primarily AI-assisted rather than fully autonomous, although it saw signs that more autonomous contributions could become more important in the near future.
Source: METR ↗ · Source: METR ↗ · Source: Anthropic ↗ · Source: Anthropic ↗
Why does it matter?
The distinction between assisted acceleration and autonomy changes what the 1.9x number means. A research team can move faster because models draft code, explore hypotheses, summarize experiments or make debugging cheaper while humans choose tasks, evaluate results and redirect the work. That can still matter for safety: if each cycle produces more capable models, the next cycle may begin with better tools. But it is not the same claim as an AI system independently deciding how to improve itself, running an open-ended research program, or recursively rewriting its own training pipeline.
The uncertainty is structural, not a footnote. AECI measures performance on tasks that can be specified and checked, and METR warns that benchmark trends can overstate acceleration if non-AI factors or the task distribution explain more of the gains. They can also understate it if recent improvements or real-world threshold effects are stronger than the benchmark shows. Redactions make the public case harder to challenge or reproduce, while the internal-model timeline is not enough to authenticate architecture, weights, availability or capabilities. The report’s estimate is informative precisely because it is narrower than the most dramatic interpretation.
For readers deciding what to watch, the useful follow-up is not another rumor about Models 2–4. It is evidence about the human and technical bottlenecks METR could not fully disclose: which R&D tasks are delegated, how often humans reject model output, how much validation time remains, and whether assisted gains persist outside well-specified benchmarks. METR’s separate time-horizon work measures the difficulty of tasks an AI can complete at a given success rate; it should not be read as a stopwatch for uninterrupted autonomous action. The defensible conclusion is that AI-assisted research may be speeding Anthropic’s capability progress, with material uncertainty and no public proof of recursive self-improvement.
Source: METR ↗ · Source: METR ↗ · Source: Anthropic ↗ · Source: Anthropic ↗
