GPT-5.4 Thinking System Card

Source: OpenAI Date: 2026-03-05 URL: https://openai.com/index/gpt-5-4-thinking-system-card

Summary

OpenAI’s system card for GPT-5.4 Thinking — the safety documentation accompanying the March 2026 GPT-5.4 mini and nano launch that included a “Thinking” variant with extended reasoning capabilities. The system card covered capability evaluations, safety assessments specific to reasoning models (where chain-of-thought introduces additional evaluation complexity), deployment guidelines, and the results of red-teaming focused on reasoning-model-specific risks — particularly around reasoning used for planning harmful actions or for deceptive outputs.

Implications

Thinking variants require distinct safety evaluation. Reasoning models like o1, o3, and now GPT-5.4 Thinking posed safety evaluation challenges not present in standard completion models: the reasoning process itself could be problematic (planning harmful actions step-by-step) even if the final output was blocked. The GPT-5.4 Thinking system card had to evaluate the reasoning trace, not just the output.

Thread: Safety documentation cadence. Part of OpenAI’s practice of publishing system cards for each major model family. The GPT-5.4 Thinking card addressed the chain-of-thought monitorability questions OpenAI had researched in December 2025.

Watch: Whether the GPT-5.4 Thinking system card reported improvements in chain-of-thought monitorability based on the December 2025 research, and what deployment restrictions if any were applied to the Thinking variant.