OpenAI announced GPT-5.4 on 5 March 2026, describing enhanced reasoning capabilities and expanded context windows [source]. The model succeeds GPT-5.3, released in January 2026, and introduces what the provider calls "improved multi-step problem solving" alongside support for 256,000-token contexts.

According to the changelog, GPT-5.4 incorporates refinements to chain-of-thought processing and claims reduced hallucination rates in technical domains. OpenAI states the model was trained on data through October 2025 and underwent additional reinforcement learning from human feedback cycles focused on mathematical and scientific reasoning.

The announcement notes that GPT-5.4 will be available to ChatGPT Plus and Enterprise subscribers beginning 12 March 2026, with API access following in April. Pricing remains unchanged from GPT-5.3 at $0.03 per 1,000 input tokens and $0.06 per 1,000 output tokens for the base variant.

OpenAI provided benchmark comparisons showing improvements on MATH, GPQA, and HumanEval datasets, though absolute scores were not disclosed in the public announcement. The provider indicated that safety evaluations were conducted by internal teams and an unnamed third-party auditor, with results to be published in a forthcoming system card.

No breaking changes to the API are documented. Existing integrations using the `gpt-5` model identifier will continue routing to GPT-5.3 until developers explicitly update to `gpt-5.4`. The changelog does not address reported drift issues affecting GPT-5.3 in February, which users documented on community forums but which OpenAI has not publicly acknowledged.