OpenAI LifeSciBench and GPT-Rosalind Push AI Toward Real Science in 2026

OpenAI just raised the bar for science-focused AI. On June 17, 2026, the company introduced OpenAI LifeSciBench, a tough new benchmark built to measure how well AI handles real life science research. Alongside it, OpenAI showed off GPT-Rosalind, a research model that already beats its predecessor on the test. Together, the two reveal both how far AI has come and how far it still has to go.

This is not another trivia quiz for chatbots. The benchmark was written and reviewed by working scientists. It asks models to do the messy, judgment-heavy work that real researchers face every day. And the early scores are humbling.

What OpenAI LifeSciBench Actually Measures

OpenAI LifeSciBench includes 750 expert-authored tasks. They span seven research workflows and seven biological domains, from genomics to medicinal chemistry. The work was shaped by 173 scientists, most holding PhDs and industry experience in biotech or pharma.

Each task looks like a request you might hand a smart lab partner. The model gets a prompt, supporting files, and a blank space for a free-response answer. Then expert-written rubrics grade the reply. Across the benchmark, those rubrics contain more than 19,000 scoring criteria, an average of 25 per task.

The design is deliberate. Real science rarely has one clean answer. A model might reach the right conclusion but miss a key caveat. The granular rubrics catch that nuance. Roughly 79% of tasks need several reasoning steps, and over half require reading figures, tables, or sequence files rather than plain text.

GPT-Rosalind Beats GPT-5.5 on Hard Research Tasks

OpenAI tested its own models against the benchmark. The newer GPT-Rosalind lifted the overall pass rate from 25.7% to 36.1% compared with GPT-5.5. That is real progress, but it also means the best model still fails roughly two tasks in three.

The biggest gains showed up in scientific communication and translation, the bench-to-bedside work of turning lab data into clinical decisions. Translation scores jumped from 36.8% to 57.7%. On tasks that demand careful handling of uncertainty and caveats, GPT-Rosalind scored 44.8% against 29.3% for the older model.

In OpenAI’s words, the model shows “meaningful progress” on this kind of expert reasoning. Even so, the company is careful not to oversell it.

Where AI Still Falls Short

The weak spots are revealing. Models struggle most with design, optimization, and analysis tasks. They also stumble when answers must be exact, such as building a CRISPR donor or designing a precise gene sequence. GPT-Rosalind hit just 14.8% on numeric tasks and 24% on sequence or structure outputs.

Artifacts are another clear gap. When a task includes figures, PDFs, or large data files, the pass rate drops sharply, from 45.1% on text-only tasks to 28.1% on artifact-heavy ones. In short, AI can talk about science well, but it is not yet a reliable hands-on collaborator at the bench.

Key takeaways

• OpenAI LifeSciBench launched June 17, 2026, with 750 expert-written tasks across seven biology domains.
• 173 scientists built it; 453 independent experts reviewed it, with over 96% agreement on quality.
• GPT-Rosalind raised the pass rate from 25.7% to 36.1%, topping GPT-5.5.
• Models shine at communication and translation but fail at exact, artifact-heavy lab work.
• The benchmark is far from saturated, leaving plenty of room to improve.

Why This Matters for Drug Discovery

Benchmarks shape where labs trust AI. A test grounded in real workflows tells research teams exactly which jobs are safe to hand off and which still need a human. That clarity matters as more companies wire AI agents into drug pipelines.

OpenAI also paired the launch with a real-world demo. The same week, it described a near-autonomous AI chemist, built with partner Molecule.one, that improved a difficult drug-making reaction. The message is consistent: AI is becoming a genuine research aid, but it works best inside clear boundaries with expert oversight.

For a wider view of how these tools are reshaping work, see our roundup of the year’s biggest shifts in our 2026 AI trends guide, and explore more coverage in our technology section.

Frequently Asked Questions

What is OpenAI LifeSciBench?
It is a benchmark of 750 expert-authored tasks that grades how well AI models handle real life science research, not simple fact recall.

What is GPT-Rosalind?
GPT-Rosalind is OpenAI’s research-focused model that improved the benchmark pass rate to 36.1%, ahead of GPT-5.5’s 25.7%.

Can AI replace scientists now?
No. The best model still fails most tasks and is weakest at exact, hands-on lab work. It functions as an assistant, not a replacement.

Where can I read the original announcement?
OpenAI published full details and the research paper on its official site.

Want more AI breakthroughs explained simply? Subscribe to the AI Hub Global newsletter for clear updates every week.

Sources: OpenAI – Introducing LifeSciBench