Epoch measured an open-weight model producing working browser exploits at roughly $1,000 each, and reports AI compute capacity doubling about every 9 months. Both bear on security planning and capability trends.
Key takeaways · AI-distilled
In Epoch's offline sandboxAn isolated environment where AI-generated code or agent actions run without being able to touch anything real.Full definition → tests, the open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → model GLM 5.3 built working exploits for six recent Firefox versions for about $1,000 each, reverse-engineering patches that had not yet reached users. Epoch could not replicate this against Chrome.
Claude Fable 5.2 scored 174 on the Epoch Capabilities Index, above GPT-6 Astra's previous record of 169 and above the trend for reasoning models. Epoch calls the score one data point in the debate over AI accelerating AI R&D.
Epoch switched ECI scoring from a frequentist to a Bayesian method, which yields head-to-head odds: GPT-6 Astra is 72% likely to be more capable than Claude Opus 5.5. Scores for models stronger than GPT-4 moved about 1 point on average.
Epoch's trade-data estimate puts 72% of the 27.4 million H100-equivalents of AI compute delivered through June 2026 in the US and 10% in China, though China holds 18% of chips, indicating many of its chips are lower-performance.
Epoch's new inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → Performance Index estimates an Nvidia GB300 serves about 25x the agentic inference capacity of an H100 despite only 2.5x the 8-bit FLOP/s, because FLOP/s misses differences in memory and inter-chip communication.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
