HealthBench Hard

How each maker reported it

• GPT-5.4 §3.6, §5.1, §6.3, p.10, p.18–28, p.36

  The card reports safety and Preparedness evaluations, not general capability benchmarks. Most values are read from the card's labelled charts; GPT-5.4 mini values come from the card's appendix tables. GPT-5.2 Thinking's Cyber Range rate is a lower bound. The health comparison columns are labelled "GPT-5.4" and "GPT-5.2". CTF is pass@12, Tacit Knowledge cons@32.

• GPT-6 Astra Launch post (openai.com/index/gpt-6-astra); system card §6, §10.1; GPT-6.1 Sol addendum §5, §9

  Coding and reasoning rows come from the GPT-6 Astra launch post (3 Sep 2026); OpenAI's own comparison table, settings as published there. The system card reports safety and Preparedness evaluations, not general capability benchmarks. Cyber and bio scores are at maximum reasoning effort; bio scores are refusal-adjusted. HealthBench scores are length-adjusted. GPT-6.1 Sol and some GPT-5.6 Sol values come from the GPT-6.1 Sol addendum (29 Sep 2026). Unverified: openai.com blocks automated access, so the coding and reasoning figures were read from copies of the launch post, and they don't appear in the system card or developer docs.

• GPT-6.1 Sol §5, §9

  The addendum reports safety and Preparedness evaluations, not general capability benchmarks. Cyber and bio scores are at maximum reasoning effort; bio scores are refusal-adjusted. HealthBench scores are length-adjusted.

• GPT-5.6 Sol p.11–15, p.38–55

  The card reports safety and Preparedness evaluations, not general capability benchmarks. HealthBench scores are length-adjusted. Most biology and cyber results are shown only as figures; values appear here only where the text or a table states them, and several (virology, ProtocolQA, TroubleshootingBench, CTF) are given only for the top-scoring GPT-5.6 model. UK AISI and Irregular rows are external evaluations of Sol. GPT-5.5 is the 'gpt-5.5-thinking'/'gpt-5.5' column of the card; its hard-negative score is the corrected 19 Aug 2026 value. Agentic-safety rows are 0–1 scores, higher is better.

• GPT-5.6 Terra p.11–15, p.38–55

  The card reports safety and Preparedness evaluations, not general capability benchmarks. HealthBench scores are length-adjusted. Most biology and cyber results are shown only as figures; values appear here only where the text or a table states them, and several (virology, ProtocolQA, TroubleshootingBench, CTF) are given only for the top-scoring GPT-5.6 model. UK AISI and Irregular rows are external evaluations of Sol. GPT-5.5 is the 'gpt-5.5-thinking'/'gpt-5.5' column of the card; its hard-negative score is the corrected 19 Aug 2026 value. Agentic-safety rows are 0–1 scores, higher is better.

• GPT-5.6 Luna p.11–15, p.38–55

  The card reports safety and Preparedness evaluations, not general capability benchmarks. HealthBench scores are length-adjusted. Most biology and cyber results are shown only as figures; values appear here only where the text or a table states them, and several (virology, ProtocolQA, TroubleshootingBench, CTF) are given only for the top-scoring GPT-5.6 model. UK AISI and Irregular rows are external evaluations of Sol. GPT-5.5 is the 'gpt-5.5-thinking'/'gpt-5.5' column of the card; its hard-negative score is the corrected 19 Aug 2026 value. Agentic-safety rows are 0–1 scores, higher is better.

• GPT-5.5 §5, §9.1, p.12, p.24–39

  The card reports safety and Preparedness evaluations, not general capability benchmarks. Several values are read from the card's labelled charts. HealthBench scores are length-adjusted; the card's health comparison column is labelled "GPT-5.4". Biochemistry is reward@4, protein binding pass@4, DNA design pass@1, CTF pass@12, Tacit Knowledge cons@32; Internal Research Debugging is a median score.

• GPT-6 Luna Appendix A, p.120–154; GPT-6.1 Sol addendum Table 9

  The card reports safety and Preparedness evaluations, not general capability benchmarks. Cyber and bio scores are at maximum reasoning effort; HealthBench scores are length-adjusted. MentalHealthBench comes from the GPT-6.1 Sol addendum.

• GPT-6 Sol Appendix A, p.120–154

  The card reports safety and Preparedness evaluations, not general capability benchmarks. Cyber scores are at maximum reasoning effort; GPT-5.6 Sol cyber values are its maximum scores using more tokens. HealthBench scores are length-adjusted. SandboxBench is targets solved out of 22. Astra values only where the appendix states them.

• gpt-oss-120b Figures 1, 2, 4; Table 2

  OpenAI's own pass@1 results at reasoning level high. AIME and HLE with tools; Codeforces with a terminal tool. o4-mini and o3 values read from OpenAI's charts.

• gpt-oss-20b Figures 1, 2, 4; Table 2

  OpenAI's own pass@1 results at reasoning level high. AIME and HLE with tools; Codeforces with a terminal tool. o3-mini was run without tools on AIME, HLE and Codeforces, so those cells are left empty. Baseline values read from OpenAI's charts.