OpenAI logo

GPT-Rosalind

Multimodal
OpenAI

GPT-Rosalind is OpenAI's specialized life-sciences reasoning model for approved internal research, introduced in research preview on 2026-04-16 and marked generally available on 2026-09-11 through OpenAI's trusted-access program for eligible Enterprise and Business organizations. OpenAI describes it as its most capable model for life-sciences research, combining deep scientific reasoning with scientific tools and databases for multi-step workflows across target research, genomics and sequencing analysis, protein and sequence analysis, medicinal chemistry, literature synthesis, wet-lab troubleshooting, and experiment planning. OpenAI lists API pricing at $5 input, $0.50 cached input, and $25 output per million tokens with no cache-write pricing, and publishes no context-window or maximum-output limits for the model. Catalog scores are self-reported from the GPT-Rosalind-5.5 system card and the LifeSciBench preprint.

Key Specifications

Parameters
-
Context
-
Release Date
September 11, 2026
Average Score
40.9%

Timeline

Key dates in the model's history
Announcement
April 16, 2026
Release
September 11, 2026
Last Update
October 8, 2026
Today
October 17, 2026

Technical Specifications

Parameters
-
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Pricing & Availability

Input (per 1M tokens)
$5.00
Output (per 1M tokens)
$25.00
Max Input Tokens
-
Max Output Tokens
-
Supported Features
Function CallingStructured OutputCode ExecutionWeb SearchBatch InferenceFine-tuning

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
Biorisk knowledge
Biorisk knowledge cons@32 from GPT-Rosalind-5.5 System Card Table 8. • Self-reported
81.1%
DNA sequence design for transcription factor binding
DNA sequence design for transcription factor binding pass@1 from GPT-Rosalind-5.5 System Card Table 11; the system card frames the task as outperforming the Ledidi baseline on Nucleobench-derived prompts. • Self-reported
13.6%
GeneBench
GeneBench pass@1 from GPT-Rosalind-5.5 System Card Table 2; the June 3 capabilities post repeats the same 21.6% figure and notes 31% fewer tokens than GPT-5.5. • Self-reported
21.6%
Hard negative protein binding prediction
Hard negative protein binding prediction pass@4 from GPT-Rosalind-5.5 System Card Table 10. Official setup gives the model PDB structures plus computer and browser access in a long-horizon agentic setting. • Self-reported
3.1%
Labwork Bench
Labwork Bench pass@1 from GPT-Rosalind-5.5 System Card Table 4; the June 3 capabilities post repeats the same 63.2% figure and compares it to 55.8% for GPT-5.5. • Self-reported
63.2%
LifeSciBench
LifeSciBench problem-weighted mean normalized score from the official OpenAI/Tacit Labs preprint. The same paper reports a 36.1% task pass rate; the GPT-Rosalind-5.5 system card separately reports a later pass@1 table, but the existing repo benchmark id tracks normalized score. • Self-reported
57.6%
MedChemBench (Internal)
MedChemBench (Internal) pass@1 from GPT-Rosalind-5.5 System Card Table 1; the June 3 capabilities post repeats the same 27.5% figure and notes 7.2% fewer tokens than GPT-5.5. • Self-reported
27.5%
Multi-select Virology Troubleshooting
Multi-select Virology Troubleshooting pass@1 from GPT-Rosalind-5.5 System Card Table 9; multimodal, multi-select setting using SecureBio-held-out virology troubleshooting questions. • Self-reported
51.2%
ProtocolQA Open-Ended
ProtocolQA Open-Ended pass@1 from GPT-Rosalind-5.5 System Card Table 6. • Self-reported
36.4%
TroubleshootingBench
TroubleshootingBench pass@1 from GPT-Rosalind-5.5 System Card Table 7. • Self-reported
53.3%

License & Metadata

License
proprietary
Announcement Date
April 16, 2026
Last Updated
October 8, 2026

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.