Fable 5.2 scores 74.6% on agentic science benchmark, ahead of Gemini Carbon and Opus 5.5

Fable 5.2 scored 74.6% on the Agentic Scientific Research benchmark. 'Agentic' means the model plans and carries out multi-step research work on its own, instead of answering one question at a time. According to the brief note we have, the score is well ahead of Google's Gemini Carbon and of Opus 5.5. The note does not give those models' scores, the size of the gap or who ran the test.

Two details matter. First, Opus 5.5 is also an Anthropic model, so part of this 'win' is a lab beating its own earlier model. Second, the comparison with Google matches what our index already shows: Google is solid but behind on the hardest tasks.

Our reading: this is the category we care about most. Our long-term optimism depends on AI speeding up science: new drug candidates, faster experiments, research that once took a decade done in a few years. A benchmark that rewards the full research workflow is a better measure of that than a chat leaderboard. Still, getting 74.6% on a test is not a discovery. The real test is results that survive peer review and work in a lab. Until we see those, this is a promising capability, not a proven change.