The independent AI safety research organization METR published on Thursday (November 5) the first deeply audited report ever conducted inside a frontier AI company. The findings are at once concrete and unsettling: Anthropic is improving its AI models at approximately twice the rate it was improving between March 2024 and October 2025, and roughly 80% of that acceleration is attributable to the use of AI in its own research and development.
What makes the Anthropic case singular is not the number itself—the doubling of progress in a year—but the methodology behind it. METR sent 2.5 full-time equivalents into Anthropic between August 12 and October 29, 2026. The team received unprecedented access to pretraining metrics, reinforcement learning data, benchmark results for unreleased internal models, AI speedup analyses, and forecasts of future model metrics. About 15 interviews with company researchers were conducted. METR received no payment from Anthropic, though it worked on Anthropic devices for security reasons.
Two years of progress in one
The central metric of the report is the AECI (Anthropic Epoch Capabilities Index), a composite of multiple benchmarks. Before October 2025, Anthropic models gained about 15 points per year on this index. After that milestone, the rate climbed to 29 points per year—a 1.9x acceleration, with a 95% confidence interval between 1.5x and 2.8x. In other words, Anthropic achieved in one year the same benchmark progress that had taken two years between October 2023 and October 2025, a leap comparable to the evolution from Claude 2.1 to Sonnet 4.5.
What the auditors confirmed with confidence is that the acceleration did not come from an increase in training compute scaling. The responsible factor is the use of AI for R&D. One researcher interviewed by METR said, 'We are now bottlenecked only on ideas and compute, no longer bottlenecked on code.' Others reported that experiments that used to take one or two days to set up are now practically 'free' in terms of overhead.
The self-reporting freefall
One of the most revealing charts in the report tracks the evolution of self-reported productivity figures submitted by Anthropic researchers. With Claude 3.7 Sonnet, reported productivity was 1.2x. With the internal Model 4 (trained between July and October 2026), employees reported productivity 12x higher than the baseline. METR treats these assessments as 'very loose evidence,' calling them unrepresentative and prone to overestimating actual gains. But the trend is consistent: even the most conservative reports indicate significant acceleration.
The report also highlights 'signs of life'—the term used by auditors—indicating that fully autonomous AI contributions to the most important parts of the training pipeline could occur soon. While AI is still not capable of making fully autonomous and high-impact contributions to the critical components of the pipeline, maturity varies across areas, and interviewed researchers agree that the inflection point lies somewhere between Opus-4.5 and the internal Model 4.
The politics of public scrutiny
The report timing is no coincidence. Anthropic is in the week prior to formal IPO marketing, with a US$100 billion offering scheduled for the week before Thanksgiving. The S-1 had already included a warning about recursive self-improvement (RSI)—the point at which AI improves itself beyond human supervisory capacity—stating that 'we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred.'
METR considers it plausible that Anthropic RSP acceleration threshold was already crossed on July 15, 2026. The company August public report stated that acceleration was 'less than a factor of 2.' The unredacted internal version suggests a value of 1.97—virtually the same number as METR. Anthropic itself deemed it 'plausible' that the threshold had been crossed.
Redactions that weaken the audit
A point that METR emphasizes repeatedly is that the public report is significantly incomplete. Anthropic exercised unilateral redaction rights over all non-public information and used them extensively. The majority of evidence that auditors rated as most informative was removed, including forecasts of future models, breakdowns by training steps, and bottleneck analysis. One chart in the public report carries a large 'FAKE DATA' watermark.
METR acknowledges that Anthropic redacted information with the legitimate intention of preventing confidential details from accelerating other AI developers—it was not merely corporate secrecy. But the net effect is that the public relies far more on METR judgment than would be ideal. Sam McCandlish, Anthropic co-founder, told The Decoder that 'many of METR conclusions match my own,' but added that it is difficult to measure exactly how much AI is accelerating research.
Dissenting voices inside Anthropic
Anthropic own researchers are not unanimous. One estimated AI contribution at only 15% to 20% of productivity gains and distrusted colleagues self-assessments. Another considered new hires the single most important factor in recent gains. When METR factored in models from OpenAI, Google, and xAI, the industry-wide acceleration number dropped to 1.4x—possibly because Anthropic lagged behind competitors in 2025, generating a 'catch-up' effect.
What this means for the industry
Acceleration at Anthropic is not an isolated phenomenon. METR calculated that, across all frontier developers, the global acceleration rate is 1.4x—lower than Anthropic but still significant. What is more alarming is the absence of observed bottlenecks. Without limits on compute, data, or diminishing returns, there is no evidence that acceleration will decelerate over the next twelve months. If AI continues automating increasingly larger parts of R&D, and if more AI in R&D generates further acceleration, the feedback loop could become self-reinforcing.
CEO Dario Amodei has already publicly warned that 'it is time to pump the brakes' on AI development, and hired METR as external auditor himself. The paradox is striking: the company that most insists on the need to slow down AI is the one that has demonstrated, in concrete data, the greatest AI-driven acceleration. And this happens exactly as Anthropic prepares to open its capital to the market—a moment when investor and public trust is more crucial than ever.
The question is whether self-sustaining AI acceleration is inevitable or whether it can be contained by institutional governance, and whether audits like METR are enough to prevent speed from escaping control.
Sources: The Verge, METR, The Decoder
✓ Independent sources cross-checked and verified before publishing
