GPT-5.4 Thinking rolled out to all ChatGPT Plus users today, bringing reasoning-model capabilities to the mainstream consumer tier. OpenAI's adaptive thinking budget automatically selects compute based on perceived problem difficulty, allowing the model to spend 5 seconds on straightforward queries and up to 60 seconds on research or proof-writing tasks.
The model scores 93% on GPQA Diamond, Google's science benchmark, and shows strong performance on AIME competition math problems. This closes the gap with Gemini 3.1 Pro's 94.3% GPQA score, narrowing what was a two-point lead a month ago. The Extended Thinking toggle is opt-in; users can enable it to force extended reasoning on hard problems, trading latency for accuracy on questions where quick answers fail.
Adaptive budgeting is the differentiator here. Rather than a fixed token budget for thinking, OpenAI's system routes queries to a routing classifier that estimates difficulty and allocates compute accordingly. A question like "What is 2+2?" bypasses thinking entirely; a research synthesis or multi-step proof gets the full budget. This reduces user-facing latency while preserving accuracy on complex queries.
Key facts
- 01OpenAI. A key thread of reporting in this story.
- 02GPT-5. A key thread of reporting in this story.
- 03Reasoning. A key thread of reporting in this story.
Pricing remains unchanged at Plus tier rates. OpenAI has not announced GPT-5.4 Thinking availability for Pro subscribers, though internal documents suggest an enterprise pilot is underway. The model is available via the ChatGPT web interface and iOS/Android apps, with API access rolling out in the coming weeks.
“GPT-5.4 Thinking reaches 93% on GPQA Diamond, closing the gap with Gemini 3.1 Pro while keeping thinking latency under 30 seconds for most queries.”— Jaeden Schafer
The release underscores a shift in OpenAI's strategy toward runtime compute over model scaling. Rather than a new o1 variant, GPT-5.4 Thinking layers adaptive reasoning atop a smaller base model, reducing inference costs while matching or exceeding last year's o1-Pro performance on benchmarks.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
