OpenAI launches GPT-5.4 across ChatGPT, API, and Codex

What OpenAI shipped

On March 5, 2026, OpenAI introduced GPT-5.4 and positioned it as its most capable and efficient frontier model for professional work. The rollout covers multiple surfaces at once: GPT-5.4 Thinking in ChatGPT, the gpt-5.4 and gpt-5.4-pro API models, and Codex support for development workflows.

OpenAI says the model combines the best reasoning, coding, and agentic-workflow improvements from its recent releases. The company also frames GPT-5.4 as the first general-purpose model with native state-of-the-art computer-use capability, which matters because more enterprise tasks now require agents to move across browsers, apps, and internal tools rather than only generate text.

The new model also expands the context window to 1 million tokens. That gives teams more room for large codebases, long policy documents, multi-file debugging, and broader retrieval pipelines without having to split work into as many separate calls.

Why the benchmarks matter

OpenAI highlighted large jumps on several benchmarks compared with GPT-5.2. On GDPval, GPT-5.4 reached 83.0 versus 70.9 for GPT-5.2. On Toolathlon it scored 54.6 versus 46.3, and on BrowseComp it reached 82.7 versus 65.8. On OSWorld-Verified, OpenAI reported 75.0 for GPT-5.4, above the cited human baseline of 72.4 and well ahead of GPT-5.2 at 47.3.

• Reasoning: stronger performance on long-horizon, high-context tasks.
• Coding: OpenAI says GPT-5.4 carries forward industry-leading coding gains from GPT-5.3-Codex.
• Computer use: better fit for agent systems that have to act across software tools, not just answer questions.

For developers and teams, the release is significant because it tightens the loop between chat, API, and agent tooling. OpenAI also said GPT-5.2 would remain available as a legacy option in the API until June 5, 2026, giving existing deployments a transition window instead of forcing an immediate migration.