Bio Capabilities Index (SecureBio)
Indexes
SecureBio's Bio Capabilities Index applies the Epoch Capabilities Index's item-response method to 13 of its own biology benchmarks, for models with scores on at least 4, anchored to GPT-4 = 70 and Claude 3.5 Sonnet = 100. It inherits the per-test accuracies and their refusal handling, and SecureBio calls it a somewhat noisy measure. Its intervals show sensitivity to the fitted parameters rather than sampling error. Diagnostic, because Atlas imports its parts separately.
Benchmark source: Securebio