OncologyBench humbles frontier AI: best model Claude Opus 5 scores just 40% on cancer tasks
kenbwork · x · 2026-10-14
Latch introduced OncologyBench, a verifiable multi-agent benchmark of 25 multistep oncology evaluations (built on 92 component tasks) spanning six stages from research to outcomes, graded deterministically against the original biological result. Strict pass rates are low: Claude Opus 5 (Pi harness) leads at 40.0%, Anthropic models take six of the top seven spots, Kimi K3 ranks 7th at 24.0% — ahead of all GPT (best 21.3%) and Grok (best 18.7%) models. Scores are reported per model×harness, since the harness substantially changes results.
Related event: OncologyBench Launch: Best AI Models Pass Only 29-40% of Oncology Decisions(6 posts)→
More from Models
- Unverified screenshots surface of a supposed Gemini 4 Argon model — ChrisUniverse · 2026-10-14
- Third-party testing suggests Mistral 4 is heavily benchmaxxed, roughly a year behind frontier — teortaxesTex · 2026-10-14
- Dev says open source lets his Sora-style video business turn a profit where OpenAI can't — ctjlewis · 2026-10-14
- Gemini 4 Argon ships in GitHub Copilot, beating prior Gemini models on every coding benchmark — rseroter · 2026-10-14
- Gemini 4 Argon appears in the app, then refuses to talk about its own release — Cyrone007 · 2026-10-14
- AnythingLLM ditches Ollama for llama.cpp, with up to 2x faster inference — solyarisoftware · 2026-10-14