Claude Fable 5.1 on science benchmarks

Claude Fable 5.1 leads the SciUniverse benchmark at a 45.3% pass rate across 92 tasks in chemistry, biology, and materials science, ahead of GPT-5 Astra at 32.5% and Claude Opus 5 at 30.5%; it was also both the highest-performing and most cost-efficient of the three at $40.61 per task. SciUniverse was released by C5R, which built Facility-0, a physical research facility integrating scientific instruments, robotics, and software, in 12 weeks. The benchmark tests whether AI can translate scientific objectives into verifiable lab results, not just recall scientific knowledge. Multiple independent evaluations corroborate Fable 5.1's lead in science-adjacent tasks: Artificial Analysis placed it atop its Intelligence Index at 1,853 Elo with a 62.0% score on SciCode; Vals AI found it first on the RSI Index for autonomous AI research at 35.03%, roughly 21% cheaper than prior leader Claude Opus 5 at $1,481 versus $1,886; and on Terminal-Bench-Science, which tests multi-step scientific workflows including reading papers, running computations, and drawing conclusions, Fable 5.1 scored 52.6%, more than doubling predecessor Fable 5's 24.7%.