What's Behind Anthropic's 2x R&D Speedup? METR's 11-Week Audit Reveals Gains and Gaps

Anthropic's R&D Speed Doubled? METR's Assessment and Its Limits

In November 2026, independent evaluation organization METR published a report assessing Anthropic's model development velocity. The core finding: on tasks with clearly defined goals and easily quantifiable outcomes, Anthropic's rate of model capability improvement reached approximately twice its historical baseline. METR estimated that roughly 80% of this acceleration may stem from AI systems themselves. However, the report includes a prominent disclaimer: because Anthropic redacted key internal data before publication — citing the need to protect sensitive proprietary information — the result carries significant potential for both upward and downward deviation. The public version retains the headline conclusion but omits the underlying internal metrics and the specific attribution logic that would allow independent verification.

Industry Context: AI Accelerating AI Research (AI4AI)

Since 2026, leading AI laboratories — including OpenAI, Google DeepMind, and Anthropic — have begun integrating model capabilities directly into their research pipelines. This trend, often discussed under the label "AI4AI" or recursive self-improvement (RSI), has made the penetration of AI into the research workflow and its acceleration effect on next-generation model development a focal point of industry attention.

Historically, model development relied heavily on researchers' intuition and manual engineering. Today, experimental design, code generation, and training optimization are progressively being delegated to models. A September 2026 paper, "The Last AI Built by Humans," formalized this progression into five levels (L1–L5), describing a shift from AI executing human-specified optimization instructions to AI leading the rewrite of subsequent improvement mechanisms, with the boundary of human intervention continuously shrinking.

Disclosed Practices at Top Labs

OpenAI ("Research acceleration," September 2026): As of mid-August 2026, each human workday invested by its research organization was matched by 3.1 workdays of agent runtime.

Google DeepMind (AlphaEvolve): Applied to a matrix-multiplication kernel split in Gemini training, yielding a 23% kernel speedup and reducing overall Gemini training time by approximately 1%.

Anthropic ("When AI builds itself"): By May 2026, over 80% of code merged into the company's codebase was generated by Claude, up from single-digit percentages before the February 2025 research preview of Claude Code.

These disclosures illustrate a broader transition: AI is moving from a "coding assistant" to an indispensable "junior researcher" and "automated engineering engine" within the model R&D pipeline. The resulting AI-driven flywheel is widening the efficiency gap between frontier labs and ordinary software teams, raising questions about whether technological progress is approaching a singularity.

METR's 11-Week Evaluation Framework

To quantify AI's actual catalytic effect, METR conducted an 11-week deep review at Anthropic. The evaluation framework centered on three dimensions:

Marginal change in the model capability growth rate.

Actual attribution weight of AI empowerment.

Whether substantive bottlenecks hindering R&D exist in the short term.

The assessment focused on task scenarios where goals are clear and results are easily scored, enabling a more controlled measurement of speedup.

Key Findings and Caveats

METR's core calculation indicates a ~2x speedup in model capability improvement under the defined task conditions, with ~80% of the gain potentially attributable to AI. However, the organization explicitly notes that missing core materials introduce significant uncertainty, meaning the true acceleration could be notably higher or lower.

Data Redaction and Credibility Concerns

Before the report's release, Anthropic applied de-identification to certain non-public information, stating the purpose was to prevent leakage of sensitive secrets that could give competitors a reference path for accelerating their own R&D. Consequently, the public report retains the high-level conclusion but hides portions of the underlying internal data and the specific attribution logic. This redaction, combined with narrow information channels and the difficulty of independently replicating the methodology, leaves Anthropic's actual R&D efficiency an open question. The report also flags factors that could lead to overestimation or underestimation of AI's objective impact on R&D acceleration, though these are not detailed in the available excerpt.

Code example

① OpenAI 在 2026 年 9 月发布的「Research acceleration」中披露，截至 8 月中旬，其研究组织每投入 1 个人类工作日，智能体的运行时间已相当于 3.1 个工作日。[2-2]
② Google DeepMind 用「AlphaEvolve」改进了 Gemini 训练中一个矩阵乘法内核的拆分方式，使该内核的运行速度提升 23%，Gemini 的整体训练时间因此缩短约 1%。[2-3]
③ Anthropic 在「When AI builds itself」里提到，到 2026 年 5 月，公司代码库合并进去的代码有八成以上出自 Claude 之手，而 2025 年 2 月 Claude Code 推出研究预览版之前，这个比例只有个位数。[2-4]