benchmarks.bio

Therapeutics · Capabilities

TxBench-Preclinical-Pharmacology results

Agentic benchmark tasks for preclinical pharmacology analysis and decision-making.

Published score: Current V1 evaluation set (100 evaluations) · short horizon · data updated 2026-11-13 (November 13, 2026)

Compare current model-and-harness results for TxBench-Preclinical-Pharmacology. Agentic benchmark tasks for preclinical pharmacology analysis and decision-making. Current V1 covers 100 evaluations and 44 ranked configurations. Data updated 2026-11-13. The highest-ranked published configurations are Gemini 4 Carbon (Pi, 66.0%); Claude Opus 5.5* (Pi, LSVP, 65.3%); Claude Opus 5.5* (Claude Code, LSVP, 65.0%). 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.

Published TxBench-Preclinical-Pharmacology scores
RankModelHarnessProviderScoreVariant
1Gemini 4 CarbonPiGoogle66.0%
2Claude Opus 5.5*PiAnthropic65.3%LSVP
3Claude Opus 5.5*Claude CodeAnthropic65.0%LSVP
4Gemini 4 ArgonPiGoogle64.7%
5Claude Sonnet 5.5PiAnthropic64.3%
6Claude Opus 5PiAnthropic63.0%
7GPT-6 SolPiOpenAI63.0%
8Claude Sonnet 5.5Claude CodeAnthropic62.3%
9Grok 4.6PiSpaceXAI61.7%
10Gemini 4 HydrogenPiGoogle61.3%
11GPT-5.6 SolPiOpenAI61.3%
12Claude Fable 5.1*PiAnthropic59.7%LSVP
13Claude Opus 5Claude CodeAnthropic59.7%
14GPT-6.1 SolPiOpenAI59.7%
15GPT-6 AstraPiOpenAI59.3%
16GPT-6 AstraCodexOpenAI59.3%
17Claude Opus 4.8PiAnthropic57.7%
18Grok 4.6Grok BuildSpaceXAI57.0%
19GPT-6.1 TerraPiOpenAI55.7%
20Gemini 3.7 FlashPiGoogle55.3%
21GPT-5.5PiOpenAI55.3%
22GPT-6.1 SolCodexOpenAI55.0%
23Gemini 3.8 FlashPiGoogle54.7%
24Grok 4.7PiSpaceXAI54.7%
25Claude Opus 4.8Claude CodeAnthropic54.3%
26GPT-6.1 TerraCodexOpenAI50.7%
27Claude Sonnet 5PiAnthropic49.3%
28Claude Sonnet 5Claude CodeAnthropic49.3%
29DeepSeek V4.1 FlashPiDeepSeek46.7%
30Claude Opus 4.7PiAnthropic45.3%
31GPT-6 LunaPiOpenAI44.3%
32GPT-5.6 LunaCodexOpenAI43.7%
33GPT-5.6 LunaPiOpenAI43.3%
34Claude Opus 4.7Claude CodeAnthropic42.3%
35Mistral Large 4.0PiMistral42.3%
36Beam 1.0PiReflection39.7%
37Kimi K3PiMoonshot39.7%
38GLM 5.3PiZ.ai35.0%
39Claude Sonnet 4.6PiAnthropic34.3%
40Claude Sonnet 4.6Claude CodeAnthropic33.3%
41Claude Opus 5.5Claude CodeAnthropic30.7%
42Claude Opus 5.5PiAnthropic26.3%
43Nemotron 3 Ultra 550B A55BPiNVIDIA22.7%
44Nemotron 3.5 LightningPiNVIDIA10.3%

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.