Domain-Specialized Model
also called Specialized model, Vertical model
A general model repackaged for one field, with that field's refusal boundary, tools and access gate.
A model offered for one domain, such as life sciences, law or cybersecurity, where the specialisation lives in the deployment around the weights rather than necessarily in new training: a field-specific refusal boundary, scientific or legal tooling, and an approval gate for who may use it. OpenAI's GPT-6 Rosalind, released October 30, 2026, is the current reference point: it is described as a frontier reasoning model for biology, drug discovery and translational medicine, built on the same weights as GPT-6 Astra with no new training data and no additional post-training.
Rosalind is the case that separates the model from the offer. The weights are Astra's, so the capability difference is policy, not parameters: Rosalind refuses fewer dual-use biology queries than Astra (9.4% versus 29.2% on OpenAI's biorisk knowledge test) and is available only through trusted access, where applicants are reviewed for a legitimate life sciences mission, a credible use case, governance and the ability to control access. Approved organisations must restrict use to authorised staff and apply qualified human review before outputs shape experiments, publications or other consequential decisions.
The benchmark story is OpenAI's own and should be quoted that way. Against GPT-Rosalind-5.5 at the highest reasoning budget tested, OpenAI reports 49.6% versus 28.9% on MedChemBench v7, 37.7% versus 16.7% on GeneBench Pro v13, and 61.3% versus 44.0% on LifeSciBench Gold. OpenAI rates the launch High capability in the biological and chemical domain, below its Critical threshold, and notes that none of its four Critical capability evaluations met that threshold.
The pricing tells the same story as the access gate. On OpenAI's pricing page as checked November 1, 2026, gpt-6-rosalind is listed at $15 per million input tokens and $75 per million output tokens, against $10 and $50 for GPT-6 Astra, the same weights. The premium buys the looser dual-use refusal boundary and the trusted-access deployment, not new parameters. That makes the buying question a governance question: is the organisation willing to maintain identity and access management, least-privilege access, administrator controls and removal processes that the addendum requires of approved customers?
The same weights, two offers
A lab is offered a specialised model at a premium over the general model it already uses.
Before paying, ask the vendor to name what changed. If the answer is the refusal boundary and the access gate, price the offer against the cost of your own governance, not against a capability gain. If the answer is new training data or post-training, ask for the training disclosure and the benchmark predecessor, because a specialised name alone proves neither.
Assuming a specialised name means specialised training, or quoting its benchmark gains without naming the predecessor and the vendor. Ask what changed: weights, refusal boundary, tools, or just the gate.
- reportOpenAI GPT-6 Rosalind: biology and drug discovery model on GPT-6 Astra weights ↗
Detailed reading of the October 30, 2026 GPT-Rosalind-6 system card addendum: same weights as GPT-6 Astra, no new training or post-training, trusted-access rules, and the three benchmark comparisons against GPT-Rosalind-5.5. All figures are OpenAI's own.
- reportGPT-Rosalind-6 brings OpenAI biology model to approved labs ↗
Access and refusal details: approved scientists, institutes, companies and government partners only; fewer refusals on dual-use biology (9.4% versus 29.2% for Astra on the biorisk test); High biological and chemical capability, below Critical.
Reasoning Model · Thinking model
A model trained to spend tokens thinking before it answers.
Foundation Model · Base model
A large model trained broadly, meant to be adapted to many downstream tasks.
Refusal
The model declining a request.
Human Review
A person reading the work, with judgment as part of the job.
Least Privilege
Give the agent the minimum access the task requires.