METR Finds Anthropic AI Research Running About 2x Faster, With AI Driving 80% of Speedup
Nonprofit evaluator METR spent 11 weeks inside Anthropic reviewing internal data, unreleased models and staff interviews, concluding the company packed about two years of model progress into the last twelve months. METR estimates benchmark progress is running 1.9x faster since October 2025, with roughly 80% of the acceleration attributed to researchers using Claude, though it calls that figure highly uncertain. Anthropic's internal Model 4 scored 178.6 on METR's AECI benchmark, and staff Slack polls reported speedups rising from 1.2x in early 2025 to 12x in October. METR acknowledged Claude performed much of its own analysis and that Anthropic redacted key evidence.
IllustrationEditorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
METR Audit Finds Anthropic AI Speeds Its Own Research About 2x, With AI Doing Most Checking
An outside group, the nonprofit METR, spent 11 weeks inside Anthropic with access to internal data, unreleased models and staff interviews to test whether Anthropic's AI is helping build its next AI. METR concluded Anthropic packed about two years of progress into the last twelve months, ruling out extra chip purchases as the main cause and estimating roughly 80% of the speedup comes from researchers using Claude. METR also acknowledged that Claude helped assemble most of the data Anthropic provided, wrote much of the documentation METR read, and performed most of METR's own analysis, including sorting interview notes behind the 2x claim. METR said cross-checking against human interviews found no bias, and Anthropic redacted much of the most convincing evidence. Internal Slack polls cited by METR rose from an average 1.2x speedup in early 2025 to 12x in October with an unreleased model called Model 4.
Read sourceMETR Report Finds Anthropic's Recursive Self-Improvement Loop Running at About 2x
A post summarizing METR's latest report states that recursive self-improvement at Anthropic has moved from theory to measurement, with the early loop running at roughly 2x. After 11 weeks reviewing Anthropic's internal R&D data and interviewing key researchers, METR concluded Anthropic is improving its models at about twice its 2024 to 2025 pace. METR's best guess attributes about 80% of current acceleration to Anthropic's own AI use, ruling out faster training compute growth as the cause. The post says Anthropic's public August risk report understated the pace, and that METR saw nothing likely to slow the loop within 12 months, with no bottleneck in algorithms, compute, or data. Redacted evidence hints acceleration may exceed 2x. Staff reportedly self-report roughly 12x productivity gains from Model 4, and Anthropic redacted most of the evidence METR found most telling.
Read sourceAnthropic's Model 4 Scores 178.6 on METR AECI, Up 4.5 Points From Model 2
Anthropic's internal Model 4 scored 178.6 on METR's AECI benchmark, roughly 4.5 points higher than Model 2's approximately 174 on the same chart, according to a post by ChrisGPT. Model 2 finished training in July, while Model 4 finished in September or October. The post also states that METR estimates benchmark progress is running roughly 1.9x faster than its previous rate, which the author frames as signs of a recursive self-improvement feedback loop. Additionally, Anthropic staff using Model 4 estimated roughly 12x their output without AI, compared with around 4x in an earlier poll using Mythos Preview. The figures are presented as reported estimates rather than independently verified results.
Read sourceShow 2 older updatesHide older updates
METR Raises AI Acceleration Estimate to 1.9x Using Anthropic Internal Models
Anthropic possesses unreleased models that were included in an independent assessment of AI progress, according to a post citing METR. METR said benchmark data from those newer internal models pushed its estimated acceleration rate to 1.9x, compared with the 1.7x reported in Anthropic's Haiku 5.5 system card. The newest model in METR's chart is labeled "Model 4" and was available internally in October 2026. The post does not detail the methodology behind the acceleration estimate or specify which models were assessed beyond the "Model 4" label.
Read sourceMETR Estimates Anthropic Model Progress Has Accelerated 1.9x Since October 2025
Independent evaluator METR estimates that Anthropic is currently advancing the capabilities of its models roughly twice as fast as it previously did, with benchmark scores improving 1.9x faster since October 2025. METR attributes a large share of this acceleration to the use of AI itself, estimating that AI may account for around 80% of the current acceleration, though it notes that this figure is highly uncertain.