GPT-5.4 is here with Computer Use, Thinking, and Pro.

Already available in Codex GPT-5.4 is here!

GPT-5.4 brings together OpenAI's recent work on reasoning, coding, and agent workflows in one model. New features include native computer use in Codex and the API, a 1 million token context window, and a tool search that helps agents pick the right tools in large MCP setups without loading everything at once. It keeps GPT-5.3-Codex's coding strengths and adds better support for spreadsheets, presentations, and documents. ChatGPT now shows you its plan before it finishes a response, so you can adjust as it goes instead of waiting for the full output.

OpenAI says GPT-5.4 is their most token-efficient reasoning model so far, using fewer tokens than GPT-5.2 for the same tasks. Benchmarks show improvements: OSWorld-Verified went from 47.3% to 75.0%, BrowseComp from 65.8% to 82.7%, and GDPval reached 83.0% compared to 70.9% for earlier models. SWE-Bench Pro is closer at 57.7% versus 56.8%, so coding improvements are smaller there. Pricing is $2.50 or $15 per million tokens, which keeps it competitive. If you're building agent workflows, it's worth comparing with Claude Opus 4.6 and Gemini 3.1 Pro. The differences between providers are getting smaller.

Source: https://openai.com/index/introducing-gpt-5-4/