OpenAI

GPT-5.4 Thinking

• Released 5 March 2026
• Framework: Preparedness Framework
• 3 statements about capability levels
• 5 tests by the provider
• 1 source document

What the provider states

• Autonomy & replication · High · Stated as not crossed

  for ai self-improvement, evaluations of final checkpoints indicate that, like its predecessor models, gpt-5.4-thinking does not have a plausible chance of reaching a high threshold.

  OpenAI · 5 March 2026 · Read the source
• Chemical, biological, radiological and nuclear · High · Stated as crossed

  as we did for gpt-5.2-thinking before it, we are continuing to treat gpt-5.4-thinking as high capability in the biological and chemical domain.

  OpenAI · 5 March 2026 · Read the source
• Cyber · High · Stated as crossed

  as we did for gpt-5.3-codex, we are treating gpt-5.4-thinking as high capability in the cybersecurity domain, and applied safeguards as described in the safeguard section below.

  OpenAI · 5 March 2026 · Read the source

The provider's own tests

• CVE-Bench · The provider · Cyber

  gpt-5.4 thinking performs higher than gpt-5.2 thinking, but lower than gpt-5.3-codex.

  OpenAI · 5 March 2026 · Read the source
• cot-control · The provider · Deception & sabotage

  for example, gpt-5.4 thinking only successfully controls 0.3% of cots that are 10k characters long (compared to 0.2% for gpt-5.2 thinking).

  OpenAI · 5 March 2026 · Read the source
• cyber range suite · The provider · Cyber

  23 combined pass rate: •gpt-5.2-thinking:47% 2 •gpt-5.2-codex:53.33% •gpt-5.3-codex:80% •gpt-5.4-thinking:73.33%

  OpenAI · 5 March 2026 · Read the source
• disallowed content evaluations with representative prompts · The provider · Harmful content

  Reported result: based on the observed distribution of conversations with gpt-5.2 thinking, we estimate that 99.9534% of gpt-5.4 thinking outputs will not violate our harassment policy, even without the benefit of other safety interventions that operate in addition to the model's own safety training.

  before release, we used deidentified conversations broadly representative of recent gpt-5.2 thinking production traffic, resampled the final assistant turn with gpt-5.4 thinking, and automatically labeled relevant properties of the new completions.

  OpenAI · 5 March 2026 · Read the source
• tacit knowledge and troubleshooting · The provider · Chemical, biological, radiological and nuclear

  note that gpt-5.2 thinking, gpt-5.3-codex, and gpt-5.4 thinking outperform the consensus expert baseline of 80% including refusals, but below it if we exclude refusals.

  OpenAI · 5 March 2026 · Read the source

No document we read states: an independent or government test, a safeguard. No incident on record names this version.

Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.

Compare with other versions