An independent evaluation has, for the first time, put a number on the feedback loop the AI industry both fears and chases: artificial intelligence helping build better artificial intelligence. The nonprofit Model Evaluation and Threat Research (METR) spent eleven weeks inside Anthropic and concluded that by October 2026 the company was advancing its models at roughly twice the pace it managed between March 2024 and October 2025 — with the majority of that speedup driven by AI systems used in research rather than by extra compute.
- Roughly 2x acceleration: METR estimates Anthropic’s progress on well-specified AI development tasks is about double its earlier rate, with a plausible range of 1.3x to 3x.
- AI as the main engine: Around 80% of the current acceleration is attributed to AI assisting research, not to greater computing power.
- Redactions limit the case: Anthropic withheld most of the evidence METR found most informative, including breakdowns of compute, data and algorithmic progress.
What METR actually measured
The assessment ran from 12 August to 29 October 2026, with the equivalent of 2.5 full-time researchers granted what METR described as unprecedented access to internal data on Anthropic’s AI research and development process. That included metrics on pretraining, reinforcement learning, training compute by domain, benchmark scores for internal models, AI speedup analyses and staff interviews.
On Anthropic’s version of the Epoch Capabilities Index, benchmark performance improved at an estimated 1.9 times its pre-October 2025 rate, with a 95% confidence interval of 1.5x to 2.8x. The broader judgment of roughly double the historical pace incorporates private evidence METR is not able to publish in full.
The human picture behind the numbers is striking. Researchers described using AI to turn ideas into experiments far faster and to explore approaches they previously lacked time to try. One researcher quoted in METR’s interview notes said the work was “now only bottlenecked on ideas & compute, no longer bottlenecked on code.” Another compared the experience to supervising postdocs and graduate students on a larger project.
The caveats are as important as the headline
METR is explicit that fully autonomous AI contributions to the most important parts of training are not yet significant, though it reports “signs of life.” The gains today come primarily from AI assisting human researchers, not replacing them.
The report also carries unusually serious transparency warnings. METR states that Anthropic redacted most of the evidence its team found most informative for its conclusions, including detailed estimates of algorithmic progress and analysis of specific constraints on future development. Anthropic has said the redactions remove confidential information that could accelerate other AI developers; co-founder and chief architect Sam McCandlish acknowledged they make it harder for outsiders to check METR’s work, but argued that publishing exactly what is driving the speedup is commercially sensitive and would amount to handing out a roadmap.
“AI is speeding up our research, but it’s hard to say exactly how much.” — Sam McCandlish, Anthropic co-founder
Other caveats pile up. Anthropic staff did not agree unanimously on the 80% attribution, and METR concedes that hiring and other non-AI factors explain much of the acceleration seen in late 2025. The company also cancelled a planned component examining how it allocates compute between capabilities and safety, and METR could not consult Anthropic’s unreleased models directly. Informal staff surveys reported productivity gains of around twelve times, a figure METR itself treats as unreliable.
Why the finding matters beyond one lab
METR found no convincing bottlenecks in compute, data or algorithms that would prevent further rapid progress over the next twelve months, though it admits its view shifted repeatedly during the assessment and may still be missing something. That matters because the speed at which capabilities compound determines how quickly advanced-model risks grow — and most of that research happens behind lab walls, out of sight of policymakers and the public.
The timing is pointed. The assessment period overlapped with Anthropic CEO Dario Amodei publicly urging the industry to slow the pace of capability improvements so safety measures can catch up, and it precedes the company’s planned initial public offering. METR also reviewed both the public and private versions of Anthropic’s August risk report and concluded the public document overall understates the level of acceleration underway.
The honest takeaway is narrower than the headline. AI is measurably accelerating AI research, the effect is real, and the exact magnitude — along with what is actually happening inside training — remains in the censored portion of the report. When the company being evaluated controls what can be published about its own evaluation, the result informs less than it appears to. What can be verified is significant enough: the loop is no longer theoretical, and the institutions measuring it are already arguing about how fast it is spinning.