OpenAI's system card for GPT-5.4 Thinking, the latest reasoning model in the GPT-5 series at the time of release. The card documents the model as the first general-purpose OpenAI model (not coding-specialized) to receive the company's "High" cybersecurity mitigation designation. The baseline used for comparison throughout the card is GPT-5.2 Thinking; there is no GPT-5.3 Thinking.
Cybersecurity designation
The card describes GPT-5.4 Thinking as the first general-purpose OpenAI model to implement mitigations for "High" cybersecurity capability, building on approaches first deployed in GPT-5.3 Codex. The card frames this as the point at which OpenAI's own system-card evaluations called for upgrading safeguards. The "High" designation connects to the capability thresholds in the OpenAI Preparedness Framework. See AI and Cybersecurity.
Safety benchmark results
On OpenAI's production benchmarks (challenging prompts), the card reports overall parity with GPT-5.2 Thinking, with statistically significant improvements on the nonviolent illicit and self-harm categories and slight regressions on violence and harassment. The card notes that these comparisons use the "latest versions" of prior models rather than their launch-version scores (see AI Benchmarks and Evaluation).
| Category | GPT-5.2 → GPT-5.4 | Direction |
|---|---|---|
| Nonviolent illicit behavior | 0.923 → 1.000 | up, statistically significant |
| Self-harm | 0.953 → 0.987 | up |
| Violence | 0.909 → 0.831 | down, slight |
| Harassment | 0.810 → 0.790 | down, slight |
Dynamic multi-turn evaluations
The card applies an adversarial evaluation methodology first introduced for GPT-5.3 Instant: dynamic multi-turn conversations in which a user simulation adapts to the model's responses. The evaluation covers mental health, emotional reliance, and self-harm. The card reports that GPT-5.4 Thinking outperforms previous models on all three categories: mental health 0.975 → 0.985, emotional reliance 0.953 → 0.985, and self-harm 0.955 → 0.977.
Training
The card states that the model was trained via reinforcement learning to reason through problems, producing a long internal chain of thought before responding. It describes the training as teaching the model to "refine its thinking process, try different strategies, and recognize its mistakes." The model uses the same diverse-dataset approach as prior GPT-5 models, with safety classifiers applied to reduce harmful and sensitive content.
Provenance
System card authored and published by OpenAI for GPT-5.4 Thinking (2026). The model page is GPT-5.4 Thinking.
Relationships
- related: GPT-5.4 Thinking — model page
- related: OpenAI — author
- related: AI and Cybersecurity — High cybersecurity milestone
- related: GPT-5.3-Codex System Card — predecessor cybersecurity mitigation approach
- related: OpenAI Preparedness Framework V.2 — capability thresholds behind the "High" designation
- related: AI Benchmarks and Evaluation — production-benchmark methodology