The safety nonprofit METR said on Thursday that Anthropic’s AI models have been improving roughly twice as fast since October 2025 as they did before, and that most of the speed-up comes from Anthropic using AI in its own research. The finding comes from an 11-week review of internal data, published in a report titled “Independent Assessment of AI R&D Acceleration at Anthropic in October 2026”. It arrives while Anthropic prepares to go public, and METR says Anthropic redacted most of the evidence it found most informative from the public version.
What METR measured
According to The Decoder’s summary of the report, a METR team of about 2.5 full-time staff had access from August 12 to October 29, 2026 to benchmark results from unreleased models, training metrics, employee surveys and internal analyses, plus around 15 interviews with Anthropic researchers. METR says it took no money from Anthropic and worked on Anthropic devices for security reasons.
The headline number is a rate of progress. Using AECI, a composite index that bundles many benchmarks, METR found Anthropic’s test scores climbing about 1.9 times faster since October 2025 than between March 2024 and October 2025. The Decoder reports that the models gained about 15 AECI points per year before October 2025 and about 29 since. The Verge quotes METR as saying this amounts to roughly two years of historical progress in the last year. The analysis includes three models Anthropic has finished training but not released, which completed training between July and October 2026.
Why METR points to AI, not compute
METR says it is confident the acceleration is not simply the result of more training compute. It estimates that the majority of the speed-up, about 80 percent according to The Decoder, traces back to AI being used in Anthropic’s research. Evidence includes interviews and informal Slack polls: employees using the most capable internal model reported being around 12 times as productive as they would be without AI, roughly double what they reported a few months earlier. METR treats those self-assessments as “very loose evidence”. One employee is quoted as saying the company is “now only bottlenecked on ideas & compute, no longer bottlenecked on code.”
For now, METR says, most of the gain comes from AI assisting human researchers rather than working independently, though it sees “some signs that fully autonomous contributions will play a larger role soon.” It also found no clear sign of diminishing returns.
The redaction problem
The report’s second story is about what the public cannot see. METR wrote that it submitted a roughly 100-page full report and that Anthropic redacted a majority of the evidence METR found most informative for the public version. METR adds that the public must now “rely on METR’s judgment much more than would be ideal.” Anthropic co-founder Sam McCandlish acknowledged on X that the redactions make it harder to check METR’s work, but said sharing what is driving the speed-up is commercially sensitive and could hand others a roadmap for going faster. He said Anthropic plans to publish its own report on the speed of AI progress in the coming weeks.
How this fits with Anthropic’s own warnings
The audit lands days after Anthropic’s amended IPO filing said the company may “imminently” meet its policy definition of dramatically accelerated progress attributable to AI, which we covered in our report on the S-1 amendment. The Decoder notes that Anthropic’s Responsible Scaling Policy treats a doubling of progress pace from automated research as a warning trigger, and its headline suggests METR’s data point to that line having been crossed as early as July. We have not seen the full report, and Anthropic has not said publicly that the trigger was met, so that reading remains METR’s and The Decoder’s interpretation.
What it means for investors and customers
The commercial backdrop matters. As we reported when Ramp’s data showed AI spending stalling, investors are already debating what Anthropic is worth, and some potential buyers reportedly see a valuation well below earlier expectations. A third-party finding that the company’s technical progress is speeding up supports one side of that debate, while the redactions and the safety implications complicate it. For customers, faster model releases mean shorter upgrade cycles and more internal models that never ship.
What to watch next
- Anthropic’s own report on the pace of AI progress, promised “in the coming weeks”.
- Any revision to the Responsible Scaling Policy that spells out how acceleration will be tracked and limited.
- Whether METR or other outside auditors get the broader “employee-like” access Anthropic described in October.
- Whether OpenAI and Google DeepMind accept comparable external reviews of their own research pipelines.
- The next S-1 amendment and the final IPO pricing.
Key takeaways
- METR says Anthropic’s model progress has run about 1.9 times faster since October 2025.
- Roughly 80 percent of the speed-up is attributed to AI used in Anthropic’s own research, per The Decoder.
- Staff self-reports of about 12 times higher productivity are weak evidence, METR cautions.
- Anthropic redacted most of the evidence METR found most informative, which limits outside checking.
- Anthropic says it will publish its own report on AI progress speed soon.
Frequently asked questions
What is METR?
METR is a nonprofit that evaluates whether advanced AI systems could pose serious risks, including whether developers could lose control of them. It says it took no payment from Anthropic for this review.
Does the report say Anthropic’s AI is improving itself without humans?
No. METR says most of the acceleration so far comes from AI assisting human researchers, though it sees signs that fully autonomous contributions may grow soon.
Why does the redaction matter?
Without the underlying evidence, outsiders cannot check METR’s conclusions and must trust its judgment. Anthropic says the redacted material is commercially sensitive and could help rivals speed up.
