OpenAI Set to Launch GPT-5.4 With 1M-Token Context Window

According to a report from The Information, OpenAI is planning the imminent release of GPT-5.4, which is to be equipped with an expanded context window of one million tokens and a new “extreme” thinking mode. The update follows the recent launches of GPT-5.3 Codex and GPT-5.3 Instant.

The new model is to feature a context window of one million tokens, more than double the 400,000 tokens of GPT-5.3. With this, OpenAI is catching up with the long-context models from Google and Anthropic. This expansion enables the processing of significantly longer documents and more complex tasks.

A key innovation is the “extreme” thinking mode. In this mode, the model applies significantly more computing time to difficult questions. OpenAI is positioning this feature primarily for scientific research and complex problem-solving, rather than for everyday users who expect quick answers.

Improved Performance on Complex Tasks

GPT-5.4 is to work more reliably on longer tasks that can take multiple hours and produce fewer errors. The model shows improved memory capabilities across multi-step workflows and lower error rates on complex tasks. These improvements are particularly relevant for OpenAI’s Codex programming tool as well as for agents and automation applications.

Background: Recent Model Releases

OpenAI has released two variants of the GPT-5.3 series in recent weeks. GPT-5.3 Codex is specifically designed for programming tasks, while GPT-5.3 Instant has been optimized for general conversations.

GPT-5.3 Instant focuses on three core areas: reduced refusals on safe requests, better integration of web information, and a more natural conversational style. According to OpenAI, the model shows 26.8 percent fewer hallucinations when using the web in critical areas such as medicine, law, and finance.

Strategy of More Frequent Updates

The release of GPT-5.4 appears to be part of OpenAI’s strategy to transition to monthly model updates. According to The Information, this approach is intended to prevent inflated expectations from building up as happened with the GPT-5 launch, which could then be disappointed. OpenAI’s user growth has recently been slower than internally predicted.

Whether GPT-5.4 can keep pace with top models from Anthropic, Google, and xAI in benchmarks and tests remains to be seen. On the platforms Arena.ai and Artificial Analysis, one can currently see that AI models from Anthropic and Google continue to rank ahead of those from OpenAI.