BioSecBench-Surveillance (PI): leaderboard
BioSecBench-Surveillance (LatchBio, 2026) gives an agent raw pathogen sequencing data and asks for a surveillance result, such as a lineage, a resistance profile or a source, graded deterministically against held-out ground truth over 100 restricted evaluations. This board is the endpoint pass rate under the PI coding agent: each evaluation's share of passing runs (three runs each, refused runs left out of the denominator), averaged over evaluations, in percent. Higher is better.
Metric: Endpoint pass rate (%). Source: github.com. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 50.2 | #56 |
| 2 | Claude Opus 4.7 | 49.6 | #75 |
| 3 | Claude Sonnet 4.6 | 48.6 | #158 |
| 4 | Gemini 3.5 Flash | 47.1 | #98 |
| 5 | Claude Opus 4.6 | 45.7 | #100 |
| 6 | Gemini 3.1 Pro (Preview) | 45.3 | #91 |
| 7 | GPT-5.5 | 44.8 | #48 |
| 8 | GPT-5.4 | 38.5 | #121 |
| 9 | Grok 4.3 | 16 | #332 |
| 10 | Grok 4.20 Beta (0309) (Reasoning) | 13.7 | #337 (Grok 4.20 Beta (0309)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=biosecbench-surveillance-pi · How It Works · Data refreshed daily, snapshot 2026-11-12.