
ALT The image shows a leaderboard titled "How good is GPT-6.2 with scientific tasks?" under "BenchBench Leaderboard: A benchmark of real-world tasks." There are three categories: Protocol, Cloning, and Refusals. In Protocol, Opus 5.3 Sol is ranked first with a 54.8% average score, followed by Gemini 4 Argon at 53.1%, and GPT-6.2 Astra at 50.6%. In Cloning, GPT-6.2 Sol is first with a 56.0% pass rate, followed by GPT-6.2 Astra at 55.8%, and Mythos 5.2 Sol at 44.6%. In Refusals, GPT-6.2 Sol is ranked first with an 85.8% Macro-F1 score, followed by GPT-6.2 Astra and GPT-6.1 Sol with 83.1%. BenchLing is mentioned at the bottom right.

ALT The image is a bar chart titled "How good are frontier models at molecular cloning? BenchBench-Cloning model performance." It shows the pass rates of different models: Mythos 5.2 LSVP (40.7%), GPT-6 Astra (30.6%), GPT-6.1 Sol (22.2%), Opus 5.5 LSVP (20.5%), and Gemini 4 Argon (18.8%). Mythos 5.2 LSVP leads, highlighted in blue. The chart includes percentage pass rates along the x-axis with an increase marker next to the overall top performer.










