openaienmodel: gpt-5-mini-2025-08-07
GPT-5.4 Thinking System Card
Key Points
- First general-purpose model with cyber safety mitigations
- Builds on GPT-5.3 Codex protections
- Baseline for comparison: GPT-5.2 Thinking
Summary
GPT-5.4 Thinking ("gpt-5.4-thinking") is the latest reasoning model in the GPT-5 series. Its safety mitigation approach follows the series precedent but is notable as the first general-purpose model to include implemented mitigations aimed at High capability cybersecurity risks. The model’s cyber safety builds on protections introduced for GPT-5.3 Codex and is integrated across ChatGPT and the API. Note: there is no GPT-5.3 Thinking model; use GPT-5.2 Thinking as the primary baseline for comparisons.
Key Points
- Review the published System Card and accompanying blog for implementation and usage guidance.
- Security: GPT-5.4 introduces targeted mitigations for high-capability cybersecurity threats—treat cybersecurity-sensitive workloads as higher risk and validate mitigations with red-team tests.
- Baseline: compare behavior and safety metrics against GPT-5.2 Thinking when evaluating regressions or improvements.
- Compatibility: protections build on GPT-5.3 Codex work; expect similar operational patterns in ChatGPT and API environments.
- Engineering actions: update threat models, run targeted security evaluations, adjust deployment gating and monitoring, and document any changes in fail-safe behavior.
- Operational note: reference the System Card for specifics before production rollout and for guidance on acceptable-use and mitigation trade-offs.