On AA's new Terminal-Bench-Science board, only two models clear half the tasks: GPT-6 Astra (max) at 63%, Claude Opus 5.5 (xhigh) at 62%. Best open-weights entry: GLM-5.3 (max) at 10%. Third-party…

Terminal-Bench-Science was made possible by our research partners @StanfordAILab, @stanfordnlp, @StanfordHAI, @2077AI, @MLFoundations, @AllenInstitute, @allen_ai, and our compute sponsors…

I submitted a science task proposal to Terminal Bench science today FYI their deadline for submitting the PR (not the merge) is Aug 17!

Most AI benchmarks test toys. This one tests whether an AI can actually do science. Terminal-Bench Science (TB-Science) is an open benchmark for AI agents that run real scientific workflows inside…

What a task looks like — real workflows from scientific domain experts: - Reconstructing MRI brain maps - Virtual drug screening - Reconstructing ice crystal disorder over time Scientifically…