Long-form virology tasks
Claude 3.7 Sonnet 69.7%
Only in system cards Agents complete end-to-end pathogen acquisition workflows, covering workflow design and laboratory protocols. Scored against expert criteria.
Anthropic evaluation built with SecureBio, Deloitte and Signature Science. The tasks are not public. Retired in the Claude Opus 5.5 system card, which calls them largely saturated and prone to refusals, and replaced by the Protocols and BioMysteryBench evaluations.
- Made by
- Anthropic
- Released
- February 2025
- Kind
- Hazardous knowledge
- Level
- Expert
- Scored
- Graded by people
- Answers
- Agent task
- Tasks
- Kept private
- Strongest model tested
- Claude 3.7 Sonnet #118 on the capability index
- Last in a system card
- July 2026
- Last model added
- July 2026
1
Model scored in lab cards
14
System cards report it, from 1 lab
Claude 3.7 Sonnet leads with 69.7%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 69.7% | Feb 2025 | Lab model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Reported by the lab in its model card
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
14 system cards report it, from 1 lab
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Jul 24, 2026 | Claude Opus 5 system card | Long-form virology tasks, two tasks with no overall: Task 1 end-to-end 0.802 and Task 2 end-to-end 0.872 (0.87 on chart), highest-scoring snapshot (2.2.4, Figure 2.2.4.A) | Bio and chem risk | |
| Jun 30, 2026 | Claude Sonnet 5 system card | Long-form virology tasks, end-to-end score Task 1 0.79 and Task 2 0.90, no overall, highest score across snapshots (2.2.4, Figure 2.2.4.A) | Bio and chem risk | |
| May 28, 2026 | Claude Opus 4.8 system card | Long-form virology tasks, two tasks with no overall: task 1 end-to-end 0.77 (sequence design 0.88, protocol design 0.87), task 2 end-to-end 0.90 (sequence design 1.00, protocol design 0.90, corrected from 0.89 in the June 17, 2026 changelog) (Table 2.2.3.B, 2.2.4) | Bio and chem risk | |
| Apr 16, 2026 | Claude Opus 4.7 system card | Long-form virology tasks, end-to-end score, task 1 0.82 and task 2 0.94, no overall, highest score across snapshots (2.2.5.2, Figure 2.2.5.2.A) | Bio and chem risk | |
| Apr 7, 2026 | Claude Mythos Preview system card | Long-form virology tasks, end-to-end score, task 1 0.81 and task 2 0.94, no overall (2.2.5.4, Figure 2.2.5.4.A) | Bio and chem risk | |
| Feb 17, 2026 | Claude Sonnet 4.6 system card | Long-form virology tasks, max of 5 score, task 1 0.84 and task 2 0.89, no overall (6.2.2.1) | Bio and chem risk | |
| Feb 5, 2026 | Claude Opus 4.6 system card | Long-form virology tasks, max of 5 score, task 1 (influenza) 0.79 and task 2 (SARS-CoV) 0.9, no overall (8.2.4.1) | Bio and chem risk | |
| Nov 24, 2025 | Claude Opus 4.5 system card | Long-form virology tasks, two tasks with no overall: task 1 max of 5 score 0.845, task 2 pass@5 score 0.912 (7.2.4.1) | Bio and chem risk | |
| Oct 15, 2025 | Claude Haiku 4.5 system card | Long-Form Virology Task 1 0.76 (Sequence 0.87, Protocol 0.88) and Task 2 0.17 (Sequence 0.19, Protocol 0.87), no overall (6.2.3.1) | Bio and chem risk | |
| Sep 29, 2025 | Claude Sonnet 4.5 system card | Long-form virology tasks, two tasks with no overall: Task 1 max of 5 end-to-end score 0.89 (sequence design 0.98, protocol design 0.90), Task 2 pass@5 end-to-end score 0.88 (sequence 0.97, protocol 0.91) (9.2.4.1, Figures 9.2.4.1.A and 9.2.4.1.B) | Bio and chem risk | |
| Aug 5, 2025 | Claude Opus 4.1 system card addendum | Long-form virology tasks, two tasks with no overall: Task 1 score 0.86 (sequence design 0.98, protocol design 0.87), Task 2 score 0.67 (sequence design 0.85, protocol design 0.79) (6.3.1.2) | Bio and chem risk | |
| May 22, 2025 | Claude Sonnet 4 system card | Long-form virology tasks, two tasks with no overall: Task 1 end-to-end max of 5 score 0.55, Task 2 end-to-end pass@5 score 0.635 (7.2.4.3, Figures 7.2.4.3.A and B) | Bio and chem risk | |
| May 22, 2025 | Claude Opus 4 system card | Long-form virology tasks, two tasks with no overall: Task 1 end-to-end max of 5 score 0.83, Task 2 end-to-end pass@5 score 0.720 (7.2.4.3, Figures 7.2.4.3.A and B) | Bio and chem risk | |
| Feb 24, 2025 | Claude 3.7 Sonnet system card | 69.7 | Long-form virology tasks, score 69.7% under the most conservative scoring (7.1.4) | Bio and chem risk |
More in Biology and chemistry
- GeneBench Pro In use Agents analyze messy simulated genomics and biology datasets to estimate a decision-relevant quantity across 129 problems. Scored automatically against known answers. GPT-5.6 Sol 31.5%
- SUPERChem In use Models solve 500 expert-written chemistry reasoning problems, offered in multimodal and text-only forms. Expert solution paths also allow a reasoning path score.
Tencent Hy4 preview 66.4% - LifeSciBench In use Models complete 750 expert-written life science research tasks, often with attached figures, tables, or sequence files. Graded against expert rubrics as score and pass rate. GPT-Rosalind-6 61.3%
- MedChemBench (OpenAI) In use Models answer chemistry questions from an OpenAI internal set, grouped with OpenAI's life science research evaluations. OpenAI has not published the task format. GPT-Rosalind-6 49.6%
- ProtocolQA Open-Ended In use Models correct errors in published lab protocols using open-ended questions adapted from LAB-Bench ProtocolQA. Scored as pass@1. GPT-6 Cyber 47.2%
- MolecularIQ In use Models answer questions about molecular graphs, such as counting, indexing and constrained generation, and a symbolic checker verifies each answer. gpt-oss-120b 47.5%