LifeSciBench: GPT-Rosalind 36.1%, artifact gap 17pts
On 2026-06-17, OpenAI published LifeSciBench — a 750-task, 1,062-artifact, 19,020-criterion life-sciences evaluation built with 173 PhD-level scientists and validated by 453 independent expert reviewers (OpenAI, Introducing LifeSciBench, 2026-06-17). GPT-Rosalind, a new life-sciences model introduced on a request-access basis, reaches 36.1% exact pass rate vs 25.7% for GPT-5.5. The under-reported finding: GPT-Rosalind drops from 45.1% on text-only tasks to 28.1% on tasks with artifacts — a 17-point drop.
What it is
LifeSciBench contains 750 expert-authored tasks across seven biological domains and seven workflows: evidence handling, analysis, design and optimization, scientific reasoning, validation and operations, translation, and scientific communication (OpenAI, Introducing LifeSciBench, 2026-06-17).
- 79% of tasks require multiple reasoning steps (avg. 4 steps/task)
- 53% of tasks require interpreting at least one artifact (figures, PDFs, sequence files)
- Grading uses per-task rubrics totaling 19,020 criteria (avg. 25/task)
- 453 independent reviewers (97% PhD, avg. 12 years experience) validated tasks; reviewer agreement exceeded 96%
The benchmark reports two metrics: pass rate (70% rubric threshold) and score (average rubric reward with partial credit). The rubric-reward number is the more honest capability signal.
The artifact-handling gap
The most useful number in the announcement is buried in the “Where AI systems still fall short” section:
- GPT-Rosalind: 45.1% text-only → 28.1% with artifacts (−17 pts)
- GPT-5.5: 29.9% text-only → 21.9% with artifacts
Both models drop the same amount. OpenAI confirms frontier models struggle at extracting information from complex figures or large sequence files (OpenAI, Introducing LifeSciBench, 2026-06-17).
For AI builders shipping agents that touch real lab data — PDFs, sequence files, gel images, instrument exports — this is the number to plan around. The 36.1% headline is a ceiling on text-only performance.
Where the gains are
- Scientific Communication: 56.3% → 71.1% — but n=9 tasks, too small to generalize
- Translation (bench-to-bedside): 36.8% → 57.7% — the strongest clean signal, on a larger category
- Design/Optimization: 30.7% — barely moved from GPT-5.5’s baseline
- Analysis: 30.3% — same story
The improvement is concentrated in talking-about workflows (communication, translation), not designing-and-analyzing workflows.
Partial-credit pattern
In roughly 14% of tasks, models earned substantial rubric credit despite failing the exact-pass threshold. GPT-Rosalind had 109 tasks with pass rates below 20% while earning at least 50% rubric reward — models can identify relevant evidence but miss key constraints.
Why it matters
The 19,020-criterion rubric is the real contribution. LifeSciBench asks “is the model useful in a research meeting?” rather than “did the model get the right answer?” The seven-workflow taxonomy maps cleanly onto how biotech R&D teams talk about their work.
Risks and caveats
- Self-report by the model owner. OpenAI is both benchmark publisher and model owner. No third-party reproduction exists as of 2026-06-20. No public code or task data is referenced in the announcement.
- 36.1% is a 70% rubric threshold, not a clean “right answer.” The rubric-reward number (44.7% vs 29.1% on expert-useful outputs) is the more honest signal.
- Scientific Communication gain is on n=9 tasks. Too small to generalize.
- Artifact handling is the load-bearing weakness. The 17-point drop is the number most useful to AI builders.
- Real research is iterative; LifeSciBench is not. OpenAI names deployment studies in live workflows as the proof point the benchmark cannot supply.
What to watch
- Public release of GPT-Rosalind access pathways (currently request-only)
- Third-party reproduction of the 36.1% / 25.7% delta
- AI Chemist and Deployment Simulation follow-ups
- Design/Optimization leaderboard updates
- Independent artifact-handling evaluations broken out by type
- Head-to-head runs against non-OpenAI models
Verdict
LifeSciBench shows frontier models improving rapidly at scientific prose and bench-to-bedside translation, and not measurably at design-and-analysis work. The 17-point artifact-handling drop is the number to plan around. It is a self-report by the model owner, GPT-Rosalind access is gated, and the strongest gain is on nine tasks (OpenAI, Introducing LifeSciBench, 2026-06-17).