← Back to feed

OpenAI’s GPT-Rosalind has exited research preview and is now available globally to eligible organizations through its trusted-access program, with published pricing in effect as of October 5, 2026.

The model launched in April 2026 under significant access restrictions — limited to US enterprise and positioned specifically around drug discovery workflows — with a BixBench score of 0.751 and third-party validation from Dyno Therapeutics — ranking in the 95th percentile against human experts on unseen RNA sequence prediction — alongside results on LABBench2. Six months later, the geographic gate is gone.

What Changed

The April launch kept Rosalind gated not just by industry but by geography, restricted to qualified US enterprise organizations. The global rollout removes that geographic restriction while keeping the trusted-access structure: organizations still apply for eligibility and access remains strictly controlled, but eligibility is no longer a US-only proposition. OpenAI says all eligible organizations will receive access to the latest Rosalind models as they’re released — a continuity guarantee not typically offered under standard API terms.

Pricing is now published: $5 per million input tokens, $0.50 per million cached input tokens, and $25 per million output tokens.

The shift from research preview to trusted-access is meaningful in OpenAI’s product taxonomy. Research preview implies internal telemetry, limited SLAs, and pricing that can change without notice. Trusted-access with published pricing means organizations can build production workflows on the model and plan for costs.

The Model Series Signal

OpenAI is explicitly framing this as “the first release in our GPT-Rosalind life sciences model series.” That language mirrors how the company has handled its general-purpose model lines — GPT-5.x, GPT-6.x — with versioned, incremental improvements. It signals Rosalind won’t be treated as a one-off specialty model but as an ongoing R&D track.

The design target is long-horizon, tool-heavy scientific workflows: multi-step experimental reasoning, literature synthesis with structured outputs, and biochemical pathway analysis. These are tasks where general-purpose frontier models run into context management and domain hallucination problems that a dedicated training corpus is intended to address.

Why Infrastructure Is the Argument

OpenAI’s stated competitive framing for Rosalind leans on compute: the ability to continuously train, evaluate, and improve against real scientific tasks as those tasks grow more complex. The implication is that the model gets better in lockstep with the difficulty of the problems being thrown at it — a feedback loop that’s harder for academic or pharma-internal AI efforts to replicate at scale.

Whether that advantage compounds in practice depends on whether the trusted-access program generates the kind of structured scientific evaluation data that actually moves benchmark needles. At launch, OpenAI featured validation from Dyno Therapeutics and results on LABBench2; the global rollout extends the pool of organizations that could surface further independent evaluations across more biology subdomains.

Key Points

• Research preview ended; global trusted-access now open to eligible organizations
• Published pricing effective October 5, 2026
• Access continuity: eligible orgs receive latest Rosalind model updates as released
• Scope: biochemical reasoning, long-horizon scientific workflows, tool-heavy tasks
• OpenAI frames Rosalind as a model series, not a standalone product