OpenAI launches GPT-5.4 with Pro and Thinking versions, 1M token context, and Tool Search

OpenAI has released GPT-5.4, which it describes as its most capable and efficient foundation model yet for professional tasks. It comes in three variants: the standard GPT-5.4, GPT-5.4 Thinking for stronger reasoning, and GPT-5.4 Pro for higher performance use cases. In the API, the model supports context windows of up to 1 million tokens, the largest OpenAI has offered so far.

OpenAI reports major benchmark gains with GPT-5.4, including record scores on OSWorld Verified, WebArena Verified, and its GDPval benchmark for knowledge work. Mercor also says GPT-5.4 leads its APEX-Agents benchmark, with strong performance on long form deliverables for law and finance. OpenAI adds that the model is more token efficient than GPT-5.2, completing similar tasks with significantly fewer tokens.

The company also says GPT-5.4 is 33% less likely to make errors in individual claims and 18% less likely to contain errors overall compared with GPT-5.2. The launch also adds Tool Search for the API, which retrieves tool definitions only when needed to reduce token use and improve speed in complex tool setups. OpenAI has also introduced a new chain of thought safety evaluation, with early results suggesting GPT-5.4 Thinking is less likely to misrepresent its reasoning. GPT-5.4 is now available in ChatGPT, Codex, and the API, with Thinking rolling out to Plus, Team, and Pro, and Pro available in the API, Edu, and Enterprise.