OpenAI released GPT-6 Rosalind on October 30, 2026, adding a specialist model option for life-sciences research teams. The API changelog describes access for authorized internal research through the Rosalind Research program.
For laboratory leaders considering a pilot, the immediate decision is about eligibility and spending: confirm the approved model, choose the product and budget the work before building an integration.
Who can use it?
Access is limited to approved Research organizations. OpenAI’s guidance specifies a limited rollout in Work and Codex, not the Chat tab. Discovery approval does not qualify, and GPT-5.5 Rosalind Research remains selectable.
Enterprise workspaces begin with a two-week Admin Preview, off by default. Administrators can enable approved Research users during it; prior approval does not bypass the preview. It is not a mandatory two-week wait.
Eligible Enterprise or Business organizations can apply through the trusted-access process. An authorized member submits the request; a verified owner or administrator completes acceptance and activation. A subscription or research-plugin installation alone does not grant access.
For software integrations, API approval is separate and must cover the organization and model, gpt-6-rosalind. Authorized internal research is permitted; customer-facing products and external commercial applications are excluded. Check activation before asking researchers to start.
What does API use cost?
The Standard API schedule lists these US-dollar rates per million tokens, the units used to meter model input and output:
Model and billing category | Input | Cached input | Output |
|---|---|---|---|
GPT-6 Rosalind: short context | $15 | $1.50 | $75 |
GPT-6 Rosalind: long context | $30 | $3 | $112.50 |
Older gpt-rosalind-research | $5 | $0.50 | $25 |
Short context covers up to 272,000 input tokens; long context is above that threshold. This billing boundary does not specify a maximum context window. Cache writes cost extra: short-context rates are $18.75 for 30 minutes and $37.50 for 12 hours per million tokens; long-context rates are $37.50 and $75.
An illustrative $30 pilot budget
Assume 100 short-context requests, each using 10,000 uncached input tokens and 2,000 billed output tokens. Applying the published GPT-6 rates gives:
100 × [(10,000 ÷ 1,000,000 × $15) + (2,000 ÷ 1,000,000 × $75)] = $30.
The same assumed token counts cost $10 at the older Research rates. This is arithmetic, not a tested workload. It excludes caching, tools, seats, taxes, discounts and extra attempts.
The threefold token-price difference is not a measured difference in cost per useful scientific result. A model that needs different token counts, more retries or more expert review could change the total substantially.
Work and Codex have their own billing terms
OpenAI’s Enterprise token-based rate card lists GPT-6 Rosalind Research at $15 input, $1.50 cached input and $75 output per million tokens in Standard mode. These rates apply to eligible token-billed agreements, not every subscription.
The card exempts Codex signed in with ChatGPT from the long-context multiplier and says Codex does not charge for cache writes. Codex using an API key follows API pricing. Confirm the sign-in route and agreement before transferring the API estimate to a workspace pilot.
For university labs, the Rosalind Academic Plan offers approved US labs up to 30 sponsored Business Standard seats for 12 months, but excludes API access and credits. Its general Rosalind benefit does not explicitly guarantee GPT-6 Research to every lab. See the academic eligibility guide for that separate application route.
What would justify paying more?
The release notice establishes a new model option, not a measured improvement in research outcomes. The access and pricing documents do not show how much researcher time GPT-6 saves or its cost per accepted scientific result.
A sensible purchase decision is therefore a bounded evaluation: confirm the approved model and product, set a spending ceiling, and define what evidence a qualified researcher needs to accept an answer. Compare with the older model where permitted, recording errors and checking time alongside token charges. Expand only if the reviewed results justify it.