The headline: Anthropic’s R&D pace may have doubled
Staff at METR (Model Evaluation and Threat Research) concluded that, by October 2026, Anthropic was likely improving the capabilities of its models on well-specified tasks at roughly twice the rate observed between March 2024 and October 2025. This acceleration suggests a significant shift in the pace of AI development.
The assessment, detailed in Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, emphasized a plausible range for this acceleration, estimating it could be between 1.3x and 3x the historical pace. Specific benchmark evaluations supported this, showing a 1.9x pace within a 95% confidence interval of 1.5x–2.8x.
This analysis focused on progress in specified tasks and internal research-and-development workflows, such as:
- AI assistance in engineering
- Automated bug fixes
- Reinforcement learning (RL) exploration
- Automated evaluation pipelines
It did not claim that every measure of AI capability is accelerating at the same rate, but rather highlighted the profound impact of AI on its own R&D cycle. Approximately 80% of this acceleration is attributed directly to AI assistance, rather than increases in pretraining compute or human capital.
The flywheel runs on AI-written code
Proposed mechanism for this acceleration centers on AI-written code. Stronger coding models empower researchers to: - Build tools - Run experiments - Debug systems - Improve training workflows
This direct assistance, in turn, expedites model development. METR’s assessment estimates AI accounts for approximately 80% of the October 2026 acceleration, though this attribution reflects a judgment, not a direct causal measurement.
A significant turning point occurred in late 2025 with the release of Claude Claude Opus. This model marked a substantial leap in AI’s ability to generate code, fundamentally altering developer workflows at Anthropic and throughout the industry.
While earlier gains from March 2024 to October 2025 were influenced by factors like increased hiring, greater compute resources, and human research breakthroughs, the post-Claude Opus period shows AI itself as the dominant accelerator. This suggests a deepening reliance on AI to drive its own progress.
A compounding loop is plausible—not inevitable
Better models could indeed increase researcher output and optimize the use of scarce compute, creating a compounding loop where gains reinforce one another. METR identified no clear near-term bottleneck that would prevent further growth, but this remains an assessment, not proof that an intelligence explosion is imminent.
A key concern is that acceleration from AI use could begin to compound, with models getting better at discovering how to improve themselves more quickly and efficiently utilizing limited compute capacity. This amplifies human researcher output, creating a self-reinforcing cycle.
However, the report also notes limits. Researchers still cited ideas and compute as constraints. While METR observed signs of life, it did not identify established autonomous contributions to the most consequential research directions. For a deeper dive into the findings, consult the Independent Assessment of AI R&D Acceleration at Anthropic in October 2026 - METR.
METR’s findings suggest Anthropic likely crossed its internal “early warning threshold” for automated R&D by July 2026. This indicates a potential for recursive self-improvement to accelerate development timelines significantly over the next one to two years.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The audit’s fine print matters as much as its forecast
METR’s assessment combined diverse evidence: interviews with Anthropic staff, self-reported productivity metrics, headcount changes, code volume analysis, experiment utilization, and qualitative reflections from researchers. Auditors received access to internal R&D processes, including model telemetry and chat histories, across an 11-week period from August to October 2026.
Crucially, Anthropic retained authority over what METR published and could deny access to certain information. While METR characterized its access as “near-unfettered,” some supporting details were redacted, requiring the public to rely on METR’s aggregated summaries and qualitative judgments rather than raw data.
This report offers a serious signal that AI is accelerating parts of frontier research. The estimated doubling of R&D speed, driven largely by AI-written code, points to a significant shift. However, the report’s estimates carry high variance (1.3x to 3x), and the limitations on METR’s access and publishing authority argue for scrutiny. This is evidence of acceleration, not certainty or runaway recursive self-improvement.
Frequently Asked Questions
What did METR find about Anthropic’s AI research pace?
METR staff estimated that, by October 2026, Anthropic was improving model capabilities on well-specified tasks at roughly twice its earlier rate, with substantial uncertainty around the estimate.
Does the report prove that AI is improving itself autonomously?
No. It finds evidence that AI assistance is accelerating parts of Anthropic’s R&D, but does not establish fully autonomous recursive self-improvement.
How much of the acceleration did METR attribute to AI?
The assessment estimated that AI accounted for the majority—perhaps about 80%—of the acceleration by October 2026.
How independent was the METR assessment?
METR conducted an external evaluation with extensive access, but Anthropic retained authority over what it could review and what findings could be published.