OncologyBench Launch: Best AI Models Pass Only 29-40% of Oncology Decisions
The Latch team has released OncologyBench, a verifiable multi-agent benchmark for oncology, showing that even today's strongest AI models achieve only a 29–40% pass rate on oncology decision-making tasks—clear evidence that large models remain far from reliably handling the full cancer care workflow.
Confirmed
- The benchmark, introduced by the Latch team, comprises 25 multi-step evaluations built on top of 92 component tasks, spanning six stages of the cancer journey: translational research, risk and detection, diagnosis and tumor characterization, treatment, response and resistance, and outcomes.
- Evaluation data comes from public datasets, case studies, and patient-donated data; each task is grounded in the real biological logic of the original experimental analyses and is verifiable.
- The full leaderboard is available at benchmarks.bio. Under the strict criterion (both task and decision fully correct), Claude Opus 5 (Pi framework) leads with 40.0%, followed by Claude Opus 5.5 (Claude Code) at 37.3%, with Anthropic models sweeping the top six spots.
- Kimi K3 is the highest-ranked non-Claude model.
Why it matters
- The benchmark covers the entire oncology care pathway from translational research through outcomes, filling a gap in verifiable, multi-step oncology evaluation.
- With top models passing only around 40%, it offers a sobering reference point for the real capabilities of AI in medical applications and a clear yardstick for future model iteration.
2026-10-14 ~ 2026-10-14 · 6 related posts
Primary sources
- [source] OncologyBench: Latch's 25-Task Multiagent Cancer Benchmark Tops Out at 40% Pass — kenbwork · 2026-10-14
- [source] OncologyBench Leaderboard: Anthropic Sweeps Top Six as Kimi K3 Leads the Rest at 24% — kenbwork · 2026-10-14
- OncologyBench: best agents pass only 29-40% of verifiable multistep cancer decision tasks — kenbwork · 2026-10-14
- OncologyBench: 25 Multistep Cancer Evaluations, Best Model Scores Only 40% — kenbwork · 2026-10-14