OpenAI launches GPT-5.4 with stronger reasoning, native computer use, and 1M Codex context

OpenAI said on March 5, 2026 that GPT-5.4 is rolling out across ChatGPT, the API, and Codex. In ChatGPT it appears as GPT-5.4 Thinking, while API customers can access `gpt-5.4` and `gpt-5.4-pro`. OpenAI described the release as its first mainline reasoning model that incorporates the frontier coding capabilities of GPT-5.3-Codex. That framing matters because it signals a tighter convergence between general reasoning, coding, tool use, and long-horizon agent workflows instead of treating them as separate product tracks.

The benchmark numbers are central to OpenAI’s case. The company reported that GPT-5.4 reaches 83.0% on GDPval, compared with 70.9% for GPT-5.2. It reported 57.7% on SWE-Bench Pro, 75.0% on OSWorld-Verified, and 82.7% on BrowseComp. OpenAI also emphasized professional knowledge-work tasks rather than just coding benchmarks. On an internal investment-banking spreadsheet benchmark, GPT-5.4 scored 87.3%, and human raters preferred GPT-5.4 presentations over GPT-5.2 outputs 68% of the time. OpenAI further said GPT-5.4 is its most factual model so far, with individual claims 33% less likely to be wrong and full responses 18% less likely to contain any error versus GPT-5.2 on a set of de-identified user prompts.

The product changes point directly at agent builders. OpenAI says GPT-5.4 is its first general-purpose model with native computer-use capabilities, which means the model is meant to operate across real interfaces and software systems rather than only generate plans or code. The company also introduced tool search so the model can work across larger tool ecosystems more efficiently, reducing prompt overhead and improving reliability in complex workflows. In Codex, GPT-5.4 adds experimental support for a 1M-token context window. OpenAI says requests beyond the standard 272K context limit count against usage at 2x the normal rate, but the larger window still gives developers a practical way to plan, execute, and verify longer tasks inside one model session.

OpenAI paired the launch with a GPT-5.4 Thinking system card that says this is the first general-purpose model in its lineup with mitigations for High capability in Cybersecurity. Pricing also moved up: `gpt-5.4` is listed at $2.50 per million input tokens and $15 per million output tokens, while `gpt-5.4-pro` is priced at $30 and $180 respectively. ChatGPT Plus, Team, and Pro users get GPT-5.4 Thinking immediately, with Enterprise and Edu customers able to enable early access. The larger picture is that OpenAI is using GPT-5.4 to bring ChatGPT, Codex, and the API closer to a single agent platform, with stronger coding, computer use, and enterprise workflow performance delivered through one flagship reasoning model.