OpenAI doesn’t publish a system card unless it’s serious about what it’s releasing. The GPT-5.4 Thinking system card dropped on March 5, 2026, and it gives us the clearest picture yet of how OpenAI is thinking about safety for its most capable reasoning model to date. If you’ve been tracking this model since launch, this document fills in a lot of gaps.
What the GPT-5.4 Thinking System Card Actually Covers
System cards are OpenAI’s way of showing their work. They document known risks, testing methodologies, mitigation strategies, and the behavioral boundaries baked into the model. For a reasoning-focused model like GPT-5.4 Thinking, that’s a bigger deal than usual.
Reasoning models think before they respond. They can chain together complex steps, revisit their own logic, and reach conclusions that simpler models can’t. That’s powerful. It’s also exactly the kind of capability that raises flags when you’re worried about misuse — someone trying to extract dangerous information, for example, or using the model’s reasoning ability to plan something harmful.
The system card addresses this head-on. OpenAI evaluated GPT-5.4 Thinking across its standard safety categories: CBRN (chemical, biological, radiological, nuclear) risks, cybersecurity, persuasion, and what they call “autonomy” risks — basically, how much the model might act in unexpected ways when given agentic tasks.
How the Safety Ratings Stack Up
OpenAI uses a tiered system for its risk ratings. GPT-5.4 Thinking landed at “medium” on a couple of key categories, which is consistent with where the company draws the line before pulling a model from deployment. Nothing here hit “high” — if it did, we almost certainly wouldn’t be talking about a public release.
That said, “medium” still means there are real considerations. The model’s extended reasoning makes it marginally better at tasks that require multi-step planning, which includes things OpenAI doesn’t want to enable. The card is transparent about the tradeoff: more capable reasoning is genuinely useful for legitimate work, but it also requires tighter guardrails in specific areas.
This is where the document gets interesting. OpenAI describes specific interventions — both at the training level and at inference time — designed to limit uplift in sensitive domains without neutering the model’s usefulness for everything else. The details are worth reading if you work in security or policy.
Agentic Use and the Autonomy Question
One section that stands out is the treatment of agentic behavior. GPT-5.4 Thinking is designed to handle complex, multi-step tasks — the kind of thing where a user sets a goal and the model figures out how to get there. That’s useful. It’s also a scenario where things can go sideways if the model takes actions the user didn’t intend or anticipate.
OpenAI’s approach here leans on what they call “minimal footprint” principles — the model should request only necessary permissions, prefer reversible actions, and check in with users when it hits ambiguous situations rather than just pushing through. This isn’t new philosophy for OpenAI, but seeing it spelled out in the context of a reasoning model adds some weight to it.
If you’ve been following GPT-5.4’s broader capabilities — the 1M context window, the coding improvements — it’s worth pairing that with what the system card says about agentic limits. The model is built for real work, but OpenAI is clearly thinking carefully about where autonomous action starts to look risky.
Evaluations and Red-Teaming
OpenAI says GPT-5.4 Thinking went through both internal red-teaming and external evaluations before release. Third-party testers were brought in specifically for the CBRN and cybersecurity categories, which is standard practice at this point but still matters. Independent eyes on the hard stuff.
The card also touches on alignment — how well the model follows instructions without being sycophantic or evasive. This is a harder problem for reasoning models because they can construct elaborate justifications. OpenAI says they’ve made progress here, though they’re measured about claiming it’s solved.
For more context on how this fits into OpenAI’s recent push into enterprise and education, it’s worth looking at what they’ve been doing with dedicated channels for AI adoption and their tools and certifications for schools. Safety documentation like this is part of the trust-building that enterprise and institutional customers require before they’ll commit.
GPT-5.4 Thinking is clearly OpenAI’s most scrutinized reasoning model yet, and the system card reflects that. As reasoning capabilities keep climbing, these documents are going to get more complex — and more important. I’d expect the next iteration to push some of those “medium” risk categories harder, which means the safety work described here is far from finished.
