Anthropic's Model Progress Has Doubled in Speed, AI-Assisted Research Likely the Main Driver
By Google News AI ·
METR, an independent AI model evaluation organization, published a research report on November 5 finding that the performance growth of Anthropic's AI models on tasks with easily verifiable correct answers has accelerated to roughly twice the previous pace. It estimates that the main driver of the acceleration is the use of AI in the research and development process.
METR analyzed Anthropic's internal data and interviews with researchers and found that AECI, an index combining multiple performance tests, has risen since October 2025 at about 1.9 times the earlier pace. The 95% confidence interval is 1.5 to 2.8 times, meaning progress that would previously have taken about two years was achieved in roughly one year.
METR concluded that the pace of increase in compute devoted to training cannot explain the acceleration. Since October 2025, growth in training compute has actually slowed, it said. What appears to contribute heavily instead is AI assistance for researchers' work.
Of the current roughly twofold acceleration, about 80% can be attributed to AI's own research support, with the remainder explained by growth in the number of researchers and expansion of inference compute, according to the estimate.
However, METR stated explicitly that AI has not yet reached the stage of autonomously handling key steps of model training and delivering major breakthroughs. The current situation is that researchers are doing the work with AI assistance, it said.
Note: Based on METR's published materials and interviews with people involved.
The report also touches on evaluation results for unreleased models. "Model 4," said to have been trained in September-October 2025, scored 178.6 on AECI, surpassing the publicly released "Opus 5.5" (174.5) and "Mythos 5.1" (172.5). Model 4 exceeded a 50% correct-answer rate on all 15 benchmarks it took, in contrast to Opus 4.6 nine months earlier, which fell short of a majority on several tests.
METR also listed as caveats the possibility that past progress was slower than assumed and that the tasks evaluated are skewed toward those with easily verifiable correct answers. It said it would not be surprising if the actual acceleration rate were around 1.3 times or 3 times.
Anthropic co-founder Sam McCandlish largely agreed with the view that AI is accelerating research, while explaining that some portions are kept private because of commercial confidentiality and the risk of accelerating other companies' development. Of the roughly 100-page report METR submitted, much of the forecast tables and analysis are redacted in the public version, limiting outside verification.
The market impact extends to the very structure of the AI development race. If the pace of model improvement doubles, companies will be forced to revisit the product launch timing and capital investment plans they have assumed. If the cycle in which AI accelerates research and development continues, the assumptions underlying compute allocation and staffing could also change.
METR said that for the next 12 months it sees no clear bottleneck in algorithms, compute, or data that would stop the acceleration.
If the measured pace persists, six months would amount to about a year's worth of progress. However, whether more advanced intelligence can be reached depends on whether gains extend to tasks without fixed correct answers, and future verification, including the transparency of evaluations, is awaited.
Originally published by Google News AI
