AI for Bio · Model

GPT-Rosalind

Life-sciences reasoning model offered through controlled access

Text, LLMs & agentsBig-tech labUnited StatesActive中文

Overview

NameGPT-Rosalind · aliases: GPT-Rosalind
Full nameA product name with no expanded form
Ambiguous names (not auto-linked)None
FamilyNot a family entry. The developer calls it the first model in a life-sciences series with no other members yet
OrganizationOpenAI
Organization typeBig-tech lab
CountryUnited States
Model classLife-sciences reasoning LLM
DomainText, LLMs & agents
First release · product update
Basis: Developer launch announcement source
Other milestones
  • First public evidence: · Launch announcement source
  • Official announcement: · Launch announcement source
Latest update† partly verified · product update
Basis: Developer update announcement
Exits research preview and opens globally to eligible organisations through controlled access, covering the enterprise assistant, the coding agent and an interface for internal tools.
Note: The official page returns 403 to fetching and the event was confirmed through a search summary; a further related announcement's date was not retrieved source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Launch and a product update within 12 months
OpennessCode: Closed
Weights: Closed
No public weights and access is controlled; the companion life-sciences plugin sits in a public repository
Benchmarks reported1 counted (modified protocol 1, unconfirmed 5)
Wet-lab validationNone recorded
Unverified fields† latest_update, training_data.a_public, training_data.b_private_licensed, training_data.c_own_wetlab_for_training, benchmarks_reported, other_evaluations
As of2026-11-16

What it does

A reasoning model for biology, drug discovery and translational medicine, used to weigh evidence, propose hypotheses, design experiments and carry out multi-step research tasks, with a companion plugin reaching more than fifty public multi-omics databases, literature sources and biology tools at inference time. Unlike open agents such as Biomni and Kosmos it is a base model behind controlled access, and unlike frontier general models it has undergone life-sciences post-training by undisclosed means.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 6 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
18 LAB-Bench / LABBench2Developer report, modified protocol—Self-reported11 of 15 LABBench2 tasks, qualitative only. †partly verified · evidence
14 PubMedQAUnconfirmed (not counted)—Unconfirmed sourceNo report found †disputed · no source link
23 BioASQUnconfirmed (not counted)—Unconfirmed sourceNo report was found. †disputed · no source link
29 BixBenchUnconfirmed (not counted)—Unconfirmed sourceVersion unspecified and the numbers appear only in second-hand reports †disputed · evidence
41 BioML-benchUnconfirmed (not counted)—Unconfirmed sourceNo report found; the only evidence is a site page, so it is shown as unconfirmed. †disputed · no source link
42 BioMysteryBenchUnconfirmed (not counted)—Unconfirmed sourceNo report found; the only evidence is a site page, so it is shown as unconfirmed. †disputed · no source link

Other evaluations (outside the 44)

  • Developer's own life-sciences evaluation (Own task set) Self-reported: 750 expert-written tasks with rubrics designed by outside experts and scored against them by another model from the same organisation
  • Internal capability evaluations (Comparison against baselines only) Self-reported: Compared with the organisation's earlier models across chemistry, protein understanding, phylogenetics, experiment design and tool use, reported only second-hand
  • Absent from the official boards (Public benchmark not among the 44) Third-party: Neither of the two relevant official boards lists it, so the self-reported figures cannot be checked against them

Developers report: the first life-sciences-specific reasoning model, leading among models with public scores on one bioinformatics analysis evaluation and beating their own previous model on part of a lab-protocol evaluation, with the clearest gain on molecular cloning questions.

Independent evidence: No independent evidence found. Neither relevant official board lists it and no third-party evaluation was found

Training data

Group in Figure 7: Not disclosed.

ClassValueDetail
a. Public databasesNot disclosedThe databases the plugin reaches are inference-time tools, not training data source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNot disclosed
d. Wet lab used only for validationNoNo public wet-lab validation; partners describe only applying or assessing it in their own workflows

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The developer is an AI lab and company based in San Francisco; this is the first model in its life-sciences line.
  • Funding: On 31 March 2026 it announced a $122B round at an $852B post-money valuation, led by three companies with existing shareholders continuing; the split is not officially disclosed.
  • Products and partnerships: Launch partners include two biopharmaceutical companies, a non-profit institute and a laboratory instruments company, with another pharmaceutical company added in a June 2026 update; earlier collaborations involved two pharmaceutical companies and an AI drug development company.
  • Related models: Biomni, Kosmos / Edison Analysis, Co-Scientist, Frontier general-purpose LLMs

Reading notes

  1. Both evaluations were run by the developer with version, prompts and tool settings undisclosed, and neither official board lists it, so nothing can be compared with board scores.
  2. The official text is qualitative and the figures appear only in second-hand reports that contradict each other, so they must be treated as second-hand.
  3. It leads on only part of the protocol evaluation, which reads as strength in one kind of protocol design rather than an across-the-board lead.
  4. Controlled access makes outside reproduction hard, and the in-house evaluation is scored by another model from the same organisation, so judge and subject share a source.

Caveats

  1. Neither weights nor training data are public and there is no system card.
  2. The official page returns 403 to fetching, so several points could be checked only via cache or search summaries.
  3. Four benchmark pairings rest on this site's own pages and must display as unconfirmed.
  4. There is no public wet-lab validation; partners describe only use within their own workflows.
  5. The in-house evaluation's scorer and the model under test come from the same organisation.

Sources

  1. Developer launch announcement (company blog)
  2. Developer update announcement (company blog)
  3. Media report (second-hand) (second-hand) second-hand
  4. Developer help page (official site)

Corrections

None recorded.

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.