What is GPT-Rosalind?
GPT-Rosalind is OpenAI's model series for life-sciences research, its first built for a single field. It launched on 16 April 2026 as a research preview for vetted US enterprises such as Amgen, Moderna and Thermo Fisher Scientific, was rebuilt on GPT-5.5 in June with better coding and tool use, and became generally available on 8 September through OpenAI's trusted-access programme for approved internal research. Novo Nordisk is among later partners.
It is tuned for biochemistry, genomics, protein engineering, drug discovery, quantitative biology and lab troubleshooting, and runs in ChatGPT, Codex and the API. Access requires approval: organisations must do legitimate research with public benefit, have governance and safety oversight, and use enterprise-grade security. Since 15 October, US university labs can apply for up to 30 free Business seats for a year, still subject to trusted-access approval.
The API model, gpt-rosalind-research, costs $5 per million input tokens and $25 per million output tokens, with cached input at $0.50; billing began on 5 October.
How much does GPT-Rosalind cost?
| Plan | Price | What you get |
|---|---|---|
| gpt-rosalind-research | $5 in / $25 out per 1M tokens | Cached input $0.50; no cache-write charges |
| Rosalind Academic Plan | Free | Up to 30 Business seats for 12 months, US university labs, approval required |
OpenAI API (approved organisations): gpt-rosalind-research $5 input / $25 output per 1M tokens, cached input $0.50; billing from 5 October 2026. US academic labs: up to 30 free Business seats for 12 months, subject to approval.
Is GPT-Rosalind worth it?
GPT-Rosalind's case rests on OpenAI's own benchmark. On LifeSciBench, 750 research tasks graded against expert-written rubrics, it scored 0.576 and passed 36.1% of tasks, ahead of GPT-5.5 (0.519, 25.7%) and Gemini 3.1 Pro (0.515, 23.6%). Anthropic's models were not included. On OpenAI's lab-troubleshooting test it scored 63.2% against 55.8% for GPT-5.5, using slightly fewer tokens.
The headline also shows the limits. Even the best model failed nearly two-thirds of the tasks, and its pass rate fell from 45.1% on text-only tasks to 28.1% when it had to work with attached files, which is where real lab data lives. With Dyno Therapeutics, its best of ten protein-prediction attempts ranked above the 95th percentile of human experts.
Every figure comes from benchmarks OpenAI designed and ran, though it has said parts will move to independent leaderboards. Treat it as a strong research assistant whose output needs a scientist's review, not an autonomous discoverer.
Access is the other constraint. It is built for pharmaceutical companies, biotech firms and, since October, US university labs that pass OpenAI's vetting; critics argue this favours large incumbents. At $5 in and $25 out per million tokens, it costs less than GPT-6 Astra, so for an approved team the price is not the barrier.
Where it earns its place
A biotech company's discovery team
Plan experiments and analyse genomics data
Use GPT-Rosalind in Codex or the API to draft analysis code and protocols, then have scientists check results; it scored 21.6% on OpenAI's GeneBench against 20.4% for GPT-5.5. A 100K-token analysis with a 10K-token report costs about $0.75.
A university biology lab in the US
Get frontier-model help without a large budget
Apply for the Rosalind Academic Plan for up to 30 free Business seats for 12 months, subject to trusted-access approval.
A wet lab with recurring failed experiments
Troubleshoot protocols
OpenAI reports 63.2% on its LabWorkBench against 55.8% for GPT-5.5. Give it the protocol and observations as text, where it performs best, rather than raw instrument files.
Before you commit
- Setup
- Apply for trusted access as an organisation; approved teams use it in ChatGPT, Codex or the API as gpt-rosalind-research. Organisations without an Enterprise account can use an OpenAI-managed workspace.
- Integrations
- ChatGPT, Codex and OpenAI API for approved organisations.
- Data and privacy
- OpenAI does not train on API data unless you opt in. Access requires governance and safety oversight and enterprise-grade security controls; usage is limited to approved internal research.
- Cost at scale
- $5/$25 per million tokens, cached input $0.50. 1,000 analyses of 100K input and 10K output tokens cost about $750 uncached; data-residency endpoints cost 10% more.
- Limits
- Approved organisations only. Weaker on tasks with attached files (28.1% pass rate on LifeSciBench). Not independently benchmarked yet.
What's good
- Leads GPT-5.5 and Gemini 3.1 Pro on OpenAI's 750-task LifeSciBench
- Better lab troubleshooting than GPT-5.5 with fewer tokens, per OpenAI
- $5/$25 per million tokens, below GPT-6 Astra
- Free seats for approved US academic labs
What's not
- Only for organisations approved through trusted access
- All benchmark figures are OpenAI's own
- Pass rate drops to 28.1% on tasks with attached files
- Academic plan limited to US universities
Best for
Approved pharmaceutical, biotech and academic research teams that want an AI research assistant tuned for biology and chemistry.
GPT-Rosalind alternatives
GPT-Rosalind: common questions
- Is GPT-Rosalind worth it?
- Measurably better than OpenAI's general models on its own life-science tests, but every number is OpenAI's, and access is limited to vetted organisations. We rate it 4.0 out of 5.
- How much does GPT-Rosalind cost?
- OpenAI API (approved organisations): gpt-rosalind-research $5 input / $25 output per 1M tokens, cached input $0.50; billing from 5 October 2026. US academic labs: up to 30 free Business seats for 12 months, subject to approval.
- What are the downsides of GPT-Rosalind?
- Only for organisations approved through trusted access; all benchmark figures are OpenAI's own; pass rate drops to 28.1% on tasks with attached files; academic plan limited to US universities.
- Who is GPT-Rosalind best for?
- Approved pharmaceutical, biotech and academic research teams that want an AI research assistant tuned for biology and chemistry.
- How much does GPT-Rosalind cost for a team or in production?
- $5/$25 per million tokens, cached input $0.50. 1,000 analyses of 100K input and 10K output tokens cost about $750 uncached; data-residency endpoints cost 10% more.
- Is GPT-Rosalind safe for company data?
- OpenAI does not train on API data unless you opt in. Access requires governance and safety oversight and enterprise-grade security controls; usage is limited to approved internal research.
- What are the best GPT-Rosalind alternatives?
- The closest alternatives we have reviewed are GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon.
Answers checked Oct 25, 2026.
Make GPT-Rosalind? Show this rating on your site
