benchmarks.bio

Omics · Capabilities

scBench results

Agentic analysis of real single-cell RNA-seq datasets across platforms and task categories.

Published score: Current V1 evaluation set (195 evaluations) · short horizon · data updated 2026-11-13 (November 13, 2026)

Compare current model-and-harness results for scBench. Agentic analysis of real single-cell RNA-seq datasets across platforms and task categories. Also known as SingleCellBench. Current V1 covers 195 evaluations and 42 ranked configurations. Data updated 2026-11-13. The highest-ranked published configurations are GPT-6 Astra (Pi, 63.9%); GPT-6.1 Sol (Codex, 63.6%); GPT-6 Astra (Codex, 63.4%). Scores report benchmark-specific biological task performance on a 0–100 percentage scale. See the linked paper and repository for task construction, grading and uncertainty methodology. Results describe the complete model and agent harness used for each run, and are not comparable across different benchmarks or evaluation versions.

Original study and current evaluation set

The original scBench study reports 394 evaluations. This leaderboard reports the current V1 evaluation set of 195 evaluations. These are separately scoped releases. Use the current V1 set when citing this table and the original study when citing its historical results. Score comparisons require the same evaluation version and scoring rules. Original study.

Published scBench scores
RankModelHarnessProviderScoreVariant
1GPT-6 AstraPiOpenAI63.9%
2GPT-6.1 SolCodexOpenAI63.6%
3GPT-6 AstraCodexOpenAI63.4%
4GPT-6.1 SolPiOpenAI62.9%
5Claude Opus 5.5*PiAnthropic61.9%LSVP
6Claude Opus 5.5*Claude CodeAnthropic61.5%LSVP
7GPT-6 SolPiOpenAI61.5%
8Claude Fable 5.1*PiAnthropic60.5%LSVP
9Claude Sonnet 5.5PiAnthropic60.3%
10GPT-5.6 SolCodexOpenAI59.8%
11Claude Sonnet 5.5Claude CodeAnthropic59.1%
12Gemini 4 ArgonPiGoogle59.1%
13Grok 4.6PiSpaceXAI59.1%
14GPT-5.6 SolPiOpenAI58.8%
15Grok 4.7PiSpaceXAI58.5%
16Claude Opus 5PiAnthropic58.3%
17GPT-6.1 TerraPiOpenAI58.3%
18GPT-6.1 TerraCodexOpenAI58.3%
19Claude Opus 5Claude CodeAnthropic57.6%
20Grok 4.6Grok BuildSpaceXAI57.6%
21Claude Opus 4.8PiAnthropic57.1%
22GPT-5.5PiOpenAI57.1%
23Claude Opus 5.5Claude CodeAnthropic56.4%
24Claude Opus 4.8Claude CodeAnthropic56.2%
25Claude Sonnet 5PiAnthropic55.6%
26Claude Opus 5.5PiAnthropic55.4%
27DeepSeek V4.1 FlashPiDeepSeek54.7%
28Claude Sonnet 5Claude CodeAnthropic53.9%
29GPT-6 LunaPiOpenAI53.9%
30Mistral Large 4.0PiMistral53.0%
31Claude Opus 4.7Claude CodeAnthropic52.5%
32Claude Opus 4.7PiAnthropic50.8%
33Gemini 3.7 FlashPiGoogle50.4%
34GPT-5.6 LunaPiOpenAI50.3%
35GPT-5.6 LunaCodexOpenAI48.9%
36GLM 5.3PiZ.ai48.7%
37Kimi K3PiMoonshot47.5%
38Claude Sonnet 4.6Claude CodeAnthropic41.7%
39Claude Sonnet 4.6PiAnthropic39.5%
40Nemotron 3 Ultra 550B A55BPiNVIDIA36.6%
41Gemini 3.8 FlashPiGoogle34.0%
42Nemotron 3.5 LightningPiNVIDIA22.9%

Methodology and provenance

Scores use a 0–100 percentage scale. See Paper PDF, Original publication, Repository for task construction, grading, aggregation, and uncertainty details.

LSVP: Anthropic's safety-filter-exempt evaluation variant, which produces fewer refusals on biology prompts. LSVP results are reported alongside the model's standard run and are not a replacement for it.