BixBench3 puts science agents at 48%

Edison Scientific’s BixBench3 found that GPT-5.6 Sol reproduced 48% of the graded artifacts across 20 research-scale computational-biology tasks, while an average attempt took 6.8 hours, processed 102 million tokens, and cost $43, according to the benchmark leaderboard. The benchmark measures execution of a scientist-specified plan—not novel hypothesis generation—and downstream artifacts scored lower than raw-data outputs; teams using agents for AI drug-discovery workflows should budget for reproducibility checks and human review rather than treating a long run as a finished result.