Overview
- The release unifies reasoning, coding and knowledge-work into one Frontier model and replaces GPT-5.3-Codex-Spark as the recommended option for developers.
- It is the first general OpenAI model with native desktop control, scoring 75% on OSWorld-Verified and posting strong BrowseComp results, while coding on Terminal-Bench 2.0 trails GPT-5.3-Codex.
- ChatGPT gains a user-interruptible 'Thinking' plan, available to Plus, Team and Pro users today, with GPT-5.2 Thinking set to retire on June 5, 2026.
- A new 'Tool Search' fetches tool definitions on demand and cut token use by 47% in tests, as API access launches under gpt-5.4 and gpt-5.4-pro with input/output prices of $2.50/$15 per million tokens that undercut Anthropic’s Opus 4.6.
- An experimental 1,000,000-token context is offered with cautions on high prefill latency and higher pricing above 272,000 input tokens, and OpenAI reports lower hallucination rates versus GPT-5.2.