Benchmark Radar™
Star —

LatchBio scBench

Explore score history ↗

195 single-cell RNA-seq analysis evaluations, but the full set is withheld to prevent training contamination and only six canonical examples are public, so scores cannot be independently reproduced. The published 95% intervals span about ±5 to 6.5 points, and the harness matters (GPT-5.5 scores 57.95 in mini-swe-agent and 57.78 in OpenAI Codex). The arXiv paper (2602.09063) reports an earlier task set, so its numbers are not comparable with these.

Evaluates
biology
Openness
unknown
Importer
Model reports

LatchBio scBench code

LatchBio scBench results and reported scores

Reported benchmark scores · Model reports

Reported scores: 19

ModelOrganizationReported valueReportedEvidence
Claude Opus 4.5Anthropic47.182026-11-16evidence
Claude Opus 4.6Anthropic52.652026-11-16evidence
Claude Opus 4.7Anthropic54.022026-11-16evidence
Claude Opus 4.7Anthropic55.212026-11-16evidence
Claude Opus 4.8Anthropic57.952026-11-16evidence
Claude Opus 4.8Anthropic57.442026-11-16evidence
Claude Sonnet 4.5Anthropic33.162026-11-16evidence
Claude Sonnet 4.6Anthropic50.262026-11-16evidence
Gemini 2.5 ProGoogle23.592026-11-16evidence
Gemini 3.1 Pro PreviewGoogle53.852026-11-16evidence
Gemini 3.5 FlashGoogle56.922026-11-16evidence
GPT-5.1OpenAI38.82026-11-16evidence
GPT-5.2OpenAI52.312026-11-16evidence
GPT-5.4OpenAI57.442026-11-16evidence
GPT-5.5OpenAI57.782026-11-16evidence
GPT-5.5OpenAI57.952026-11-16evidence
Grok 4.1 Fast (reasoning)xAI30.262026-11-16evidence
Grok 4.20 Beta 0309 (reasoning)xAI44.442026-11-16evidence
Grok 4.3xAI44.272026-11-16evidence

Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.

Which sources cite LatchBio scBench?

Related biology benchmarks