An external review suggests AI tools may be helping Anthropic build better models at a faster rate. METR, a nonprofit that studies risks from advanced AI, estimates the company’s model scores have risen about 1.9 times faster since October 2025. It attributes roughly 80 percent of that acceleration to AI used in Anthropic’s own research, while warning that important evidence was withheld from public view.
What METR measured
METR examined Anthropic’s internal research data over eleven weeks. From August 12 to October 29, 2026, 2.5 full-time staffers reviewed benchmark results for unreleased models, training metrics, employee surveys and internal analyses. They also conducted roughly 15 interviews with Anthropic researchers.
The analysis uses the AECI, a composite measure that combines results from multiple benchmarks. METR says Anthropic’s models gained about 15 AECI points per year before October 2025, compared with 29 points per year since then. In practical terms, the nonprofit describes a year of recent progress as roughly equivalent to what had previously taken two years.
The review included three unreleased models that finished training between July and October 2026. METR compares the overall change to the leap from Claude 2.1 to Sonnet 4.5. It says the acceleration does not appear to come from training compute growing faster.
AI assistance may be speeding up research
METR’s estimate is that about 80 percent of the recent speedup stems from AI being used in Anthropic’s research. That points to a practical feedback loop: AI tools help researchers with work that contributes to better models, which can then assist with further research.
The accounts collected by METR suggest that AI is chiefly acting as an assistant rather than conducting research on its own. One employee described a shift in what limits their work: “Now only bottlenecked on ideas & compute, no longer bottlenecked on code.” Others said experiments that once took one to two days to set up had become “~free.”
In informal Slack polls from October 2026, employees estimated that the latest internal model made them roughly twelve times more productive on average. METR calls those responses “very loose evidence”: they are self-reported, may not represent the workforce and could exaggerate the gains. One researcher put AI’s contribution to productivity at just 15 to 20 percent, while another pointed to new hires as the biggest factor in recent months.
A warning threshold with a wide range
Anthropic’s Responsible Scaling Policy treats a doubling in the pace of AI progress due to automated AI research as a warning-level signal. METR’s central estimate for Anthropic is about 1.9 times faster, close to that threshold. Its companion review says the company’s August 2026 risk report publicly described acceleration as “less than a factor of 2,” while an internal, unredacted version implies 1.97. Anthropic itself called it “plausible” that the threshold had been crossed.
METR says the warning level may have been reached on July 15. That is a possibility, not a definitive finding: the nonprofit says the true acceleration could range from 1.3 to 3. When results from OpenAI, Google and xAI are included, its estimate for the wider industry drops to 1.4. METR says a catch-up effect may be involved because Anthropic lagged behind competitors in 2025.
The nonprofit found no bottlenecks that it expects to slow progress over the next twelve months, and sees a significant risk that acceleration could continue growing. At the same time, METR shares Anthropic’s assessment that acute catastrophic risk is “low.” It says another sharp acceleration over the next six months could raise that risk rating quickly.
Redactions and a missing audit
Anthropic had unilateral rights to redact non-public information and used them extensively. METR says the majority of the evidence it considered most informative was removed, including forecasts for future models, breakdowns by training steps and analysis of potential bottlenecks. Anthropic said the redactions were “aimed at removing confidential information that could accelerate other AI developers.”
With much of the supporting material unavailable, readers must depend more heavily on METR’s interpretation than the nonprofit considers ideal. One graph in the report displays “FAKE DATA” in large letters, reflecting that the underlying evidence cannot be shown openly.
A separate part of the planned review was canceled midway through: METR had been asked to check how Anthropic divides compute between capability and safety research. Anthropic said it wanted to focus first on risk assessment and revisit that audit later. Daniel Kokotajlo, the researcher behind the “AI 2027” scenario, argued on X that verifying compute allocation is foundational to making commitments to slow down verifiable.
The review provides a rare look at how AI tools may be affecting AI development inside one company. Its estimate is notable, but the wide uncertainty, differing employee views and extensive redactions leave key parts of the case difficult for outsiders to evaluate.