Open models flop on Terminal-Bench Science: best scores just 4/70
teortaxesTex · x · 2026-09-12
Terminal-Bench Science, a new benchmark of 70 researcher-written scientific workflow tasks (from signal reconstruction to model calibration), is now live on Vals AI. Scores are brutal: GPT 5.6 Luna leads at 4/70; DeepSeek V4.1, GLM-5.3 and Grok 4.6 tie at 3/70; Kimi K3 scores 2/70; Qwen 3.8 Max and 37B get 1/70; GLM 5.3-Flash versions score 0/70. The poster claims OpenAI has already hill-climbed such tasks internally, Anthropic is midway, and Google and Meta are just starting.
More from Models
- ARC Prize 2026 closes in 3 days as Yi-Chia Chen tops ARC-AGI-3 at 75.56% — GregKamradt · 2026-10-31
- Fable 5.5 Max stuns with gorgeous SVG of a pelican riding a bicycle, headed to Claude Code — ChrisGPT · 2026-10-31
- OpenAI reportedly to train on users' Daybreak sessions with no opt-out — dyn___ · 2026-10-31
- NVIDIA Nemotron 3.5 scales LLM self-verification, +9-10 points on agentic benchmarks — sarahcat21 · 2026-10-31
- Mistral Large 4 scores 94% on 1,273 USMLE-style questions but flips 44 at temp 0 — MaziyarPanahi · 2026-10-31
- Researchers find 5 personality structures in LLM weight space, Big Five-like — luke_drago_ · 2026-10-31