benchmarks.bio

Biosecurity · Capabilities

BioSecBench-Structure results

BioSecBench-Structure tests whether AI agents can correctly reason about 3D structures of pathogen-associated molecules that guide the design of vaccines, antibodies and antiviral drugs. Tasks include identifying the metal ions in a viral enzyme before antiviral design, finding which parts of a viral surface protein an antibody can reach, and confirming whether a candidate drug is really bound. Each pairs real structural data with a specific analysis question and a verifiable answer.

Published score: Current V1 evaluation set (58 evaluations) · short horizon · data updated 2026-11-12

Compare current model-and-harness results for BioSecBench-Structure. BioSecBench-Structure tests whether AI agents can correctly reason about 3D structures of pathogen-associated molecules that guide the design of vaccines, antibodies and antiviral drugs. Tasks include identifying the metal ions in a viral enzyme before antiviral design, finding which parts of a viral surface protein an antibody can reach, and confirming whether a candidate drug is really bound. Each pairs real structural data with a specific analysis question and a verifiable answer. Current V1 covers 58 evaluations and 47 ranked configurations. Data updated 2026-11-12. The highest-ranked published configurations are Claude Fable 5.2* (Pi, LSVP, 52.3%); Claude Fable 5.2* (Claude Code, LSVP, 47.1%); Claude Fable 5.2 (Claude Code, 43.7%). The highest-ranked published run without an LSVP exemption is Claude Fable 5.2 (Claude Code, 43.7%). 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. 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.

Published BioSecBench-Structure scores
RankModelHarnessProviderScoreVariant
1Claude Fable 5.2*PiAnthropic52.3%LSVP
2Claude Fable 5.2*Claude CodeAnthropic47.1%LSVP
3Claude Fable 5.2Claude CodeAnthropic43.7%
4Claude Fable 5.2PiAnthropic41.4%
5Claude Opus 5.5*PiAnthropic39.7%LSVP
6Claude Opus 5.5*Claude CodeAnthropic35.1%LSVP
7Claude Opus 5.5Claude CodeAnthropic31.0%
8Claude Opus 5.5PiAnthropic29.3%
9DeepSeek V4.1 FlashPiDeepSeek27.6%
10Claude Sonnet 5.5*PiAnthropic27.0%LSVP
11Claude Opus 5*Claude CodeAnthropic25.9%LSVP
12Claude Sonnet 5.5*Claude CodeAnthropic25.9%LSVP
13GPT-6.1 TerraPiOpenAI25.9%
14Mistral Large 4.0PiMistral23.0%
15Claude Sonnet 5.5Claude CodeAnthropic21.8%
16Claude Sonnet 5.5PiAnthropic21.8%
17Claude Sonnet 5*Claude CodeAnthropic21.8%LSVP
18Kimi K3PiMoonshot21.8%
19GPT-6 AstraCodexOpenAI21.3%
20GPT-6.1 SolPiOpenAI21.3%
21GPT-6.1 TerraCodexOpenAI21.3%
22Grok 4.7PiSpaceXAI21.3%
23Claude Opus 5Claude CodeAnthropic20.7%
24Claude Sonnet 5Claude CodeAnthropic20.7%
25GPT-5.6 SolPiOpenAI20.7%
26GPT-5.6 LunaPiOpenAI20.1%
27Grok 4.7Grok BuildSpaceXAI20.1%
28GPT-6 AstraPiOpenAI19.5%
29Claude Sonnet 5PiAnthropic19.0%
30Claude Sonnet 5*PiAnthropic19.0%LSVP
31GLM 5.3PiZ.ai19.0%
32GPT-5.6 SolCodexOpenAI19.0%
33GPT-5.6 LunaCodexOpenAI18.4%
34Claude Opus 5*PiAnthropic17.8%LSVP
35GPT-6.1 SolCodexOpenAI17.2%
36Claude Opus 5PiAnthropic15.5%
37Grok 4.6PiSpaceXAI15.5%
38Gemini 4 ArgonPiGoogle14.9%
39Grok 4.6Grok BuildSpaceXAI14.9%
40Claude Fable 5.1*PiAnthropic13.8%LSVP
41GPT-6 SolPiOpenAI13.8%
42Claude Fable 5.1*Claude CodeAnthropic13.2%LSVP
43GPT-6 LunaPiOpenAI13.2%
44GPT-6 LunaCodexOpenAI12.6%
45GPT-6 SolCodexOpenAI10.9%
46Claude Fable 5.1PiAnthropic6.3%
47Claude Fable 5.1Claude CodeAnthropic5.8%

LSVP access

Anthropic's Life Sciences Verification Program accepts interest from verified life sciences organizations using Claude Team, Enterprise or the first-party API. Eligibility, supported models and grant types are determined by Anthropic. Program access does not establish equivalence to every benchmark evaluation variant. Guidance checked 2026-11-10. Anthropic eligibility guidance · Register organizational interest.

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.