METR Finds Anthropic's AI-Assisted R&D Moving Twice as Fast

During an 11-week assessment inside Anthropic, researchers from METR examined pretraining metrics, reinforcement-learning logs, compute budgets, internal surveys, and researcher interviews. They concluded that AI tools are probably accelerating Anthropic’s model development, though redactions and measurement limits leave substantial uncertainty.

As of October 2026, METR estimates that Anthropic was improving model performance on well-specified tasks at about twice the rate recorded from March 2024 through October 2025. Its overall uncertainty range spans roughly 1.3× to 3×. At the central estimate, Anthropic achieved about two years of its previous capability gains within 12 months.

The curve steepened

Anthropic’s internal AECI benchmark provided METR’s main quantitative evidence. AECI, Anthropic’s version of Epoch’s Capabilities Index, combines results from many benchmarks into a statistically estimated capability score.

A segmented analysis found that AECI scores rose by about 15 points per year before October 2025 and 29 points per year afterward. That produces a 1.9× acceleration estimate, with a 95% confidence interval of 1.5× to 2.8×.

METR’s alternative analyses produced a wide range, but most still showed acceleration. The lower estimate across all frontier models also suggests that part of Anthropic’s measured gain reflects catching up from a weaker position during early 2025.

AI tools explain most of the gain

METR attempted to divide the acceleration among training compute, data, algorithmic advances, hiring, and AI-assisted research. Its tentative estimate attributes about 80% of the current acceleration to AI use. The team also reported convincing evidence that faster growth in training compute did not explain the change.

Informal Slack polls supplied one measure of AI’s effect on individual employees. For Anthropic’s most capable internal model in October 2026, respondents reported an average productivity multiplier of roughly 12×, up from about 1.2× for Claude 3.7 Sonnet in early 2025.

Those informal estimates cannot be converted directly into company-wide R&D acceleration. Self-reports can also overstate measured effects: METR’s 2025 developer study found that AI tools slowed experienced open-source developers on the tasks tested.

METR cross-checked the surveys by interviewing 15 Anthropic researchers and using Claude to classify their responses. The published categories account for 11 interviewees:

• 2 researchers clearly supported a speedup of at least 2× in their area.
• 3 researchers leaned toward a speedup of at least 2×.
• 4 researchers gave borderline or unclear answers.
• 2 researchers rejected the estimate and attributed most growth to hiring or other factors.

The listed breakdown leaves four interviews unaccounted for. Among the responses METR discussed, one researcher endorsed more than two years of progress within a year and considered a fourfold rate plausible. Another attributed only 15% to 20% of the acceleration to AI. Interviewees also described early AI contributions to research work, with maturity varying by field; fully autonomous research contributions were not yet a major driver.

No 12-month bottleneck surfaced

METR found no evidence that algorithmic progress, training compute, experiment compute, or data would impose a significant constraint during the following 12 months. Continued feedback between better models and faster research could therefore raise the development rate further. METR revised its view of potential bottlenecks several times during the assessment, so this finding records the absence of a detected constraint rather than a firm forecast.

The strongest evidence remains hidden

Anthropic received METR’s full 100-page report but redacted most of the evidence that METR considered most informative. The company said the edits were aimed at removing confidential information that could accelerate other AI developers. The withheld material includes detailed estimates of algorithmic progress, attribution by research pipeline, and forecasts for future models, limiting independent scrutiny of the headline conclusions.

Limits on the estimate

• AECI is approaching saturation. The report’s “Model 4” scores above 50% on every benchmark in the index, reducing the benchmark’s ability to distinguish future gains.
• The tasks are unusually clean. AECI emphasizes work with verifiable answers. Performance on ambiguous production tasks may follow a different trajectory.
• Anthropic mediated access. METR had to request data proactively while working on company devices and could not query unreleased models directly. The team may have missed systems or evidence it did not know to request.
• The compute-allocation audit was canceled. The original scope included checking how Anthropic divided compute between capability and safety work. Anthropic ended that portion of the pilot before METR reached a conclusion.
• Much of the attribution is observational. Hiring, organizational changes, research choices, and AI adoption occurred together, making their separate effects difficult to identify.

Release cadence may tighten

Anthropic identifies AI-driven R&D acceleration as a key risk indicator in its scaling policy. METR’s assessment provides the first structured outside review of the internal evidence behind that indicator, including a benchmark methodology, uncertainty ranges, employee surveys, and researcher interviews.

For developers building on frontier models, faster internal progress could shorten the useful life of model-specific evaluations, prompts, cost assumptions, and capability forecasts. Anthropic’s unreleased “Model 4” reportedly exceeds its public systems, while the pipeline producing new models appears to be accelerating. The assessment cannot establish whether public releases will follow the same pace or whether gains on verifiable benchmarks will transfer to complex production work.