METR published an independent assessment on November 5, 2026 after its staff examined non-public information about Anthropic's AI research and development. Its central judgment is that Anthropic may now be improving model capabilities on well-specified tasks at roughly twice its earlier pace, with AI assistance contributing substantially. That is the evaluators' uncertain judgment about one company's research process, not a measured doubling of every kind of software development or a prediction that the next model will arrive twice as fast.
What did METR actually measure?
METR's team had access from August 12 to October 29 to internal benchmarks, training-compute information, researcher interviews and other research-process evidence. One public quantitative result is that scores on Anthropic's AECI benchmark series grew at an estimated 1.9 times their pre-October-2025 rate since October 2025, with a reported 95% confidence interval of 1.5 to 2.8 times. AECI is a measure of model performance on selected tasks, so this rate is not a direct measurement of human productivity or real-world app quality.
METR says the increase was not explained by faster growth in training compute. Interviews and surveys were consistent with AI helping researchers work faster, although self-reports cannot settle cause on their own. The authors' broader estimate is that a majority of the current acceleration is likely attributable to AI use; they also say non-AI factors probably mattered, especially earlier in the period. Much of the work described was AI assistance to researchers rather than fully autonomous research.
How strong is the evidence for AI causing the acceleration?
The report is unusually candid about the limits of a public assessment. METR says Anthropic redacted a majority of the evidence the evaluators found most informative for their highest-level conclusions. The public therefore cannot independently inspect all the material behind the causal judgment. The authors also discuss alternative explanations, benchmark design and sparse recent data; they say an actual acceleration closer to 1.3 or 3 times the earlier rate would not surprise them.
Anthropic had announced employee-like evaluator access in October as part of its independent assessment program. METR describes its own access, methods and redaction process in the November report. Independence of the evaluator does not erase the limits created by restricted publication, and the release should be read with both features in view. VIBPS has not audited the underlying private data.
Why does this matter to software makers?
If AI tools help a frontier lab iterate on its own models, the feedback loop could affect the tools that app makers later use to build software. That is a possible downstream consequence, not a result directly measured by this assessment. The evidence here concerns Anthropic's model R&D and tasks that can be evaluated relatively clearly; it does not show that an individual can now create, ship or grow an app twice as fast.
The practical reading is to keep measurement specific. Faster benchmark progress, researcher-reported help and public product usefulness are different claims. A maker choosing AI tools still needs to assess their own code quality, iteration time and user outcomes. VIBPS covers this research because software creation is changing, while the audience and social response to what people build remain separate questions.