Introducing GPT-5.4
OpenAI
OpenAI just dropped GPT-5.4, its new top model for pro work. It claims best-in-class coding and a million-token memory, so it can read whole codebases at once.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI has rolled out GPT-5.4, positioning it as the new default for anyone doing serious professional work rather than casual chatting. The pitch is efficiency plus raw capability, a combination OpenAI has been chasing since GPT-5's launch complicated its own lineup with too many overlapping variants.
The headline number is the 1-million-token context window, big enough to swallow entire codebases, lengthy legal contracts, or sprawling research papers in a single pass without the constant chunking and re-summarizing that older models required. Pair that with what OpenAI calls state-of-the-art coding and computer-use skills, and the target audience becomes obvious: developers, analysts, and agents that need to actually operate software, not just describe it.
Tool search is the quieter addition here, and possibly the more consequential one. Instead of a model guessing which of a dozen bolted-on tools might help, GPT-5.4 apparently searches for the right tool on the fly, a small architectural shift that matters a lot once you're running agents against hundreds of possible functions instead of five.
OpenAI is light on benchmark specifics in this announcement, favoring broad claims over hard numbers, which is typical for a first-day blog post. But the framing — efficient, frontier, built for professional workloads — signals a company still trying to convince enterprise customers that its models are worth the switching costs, especially as Anthropic and Google keep shipping credible alternatives at similar price points.
My take — AI-written commentary, not fact-checked reporting
I'll believe 'efficient frontier model' when someone publishes actual latency and cost numbers instead of a press release. OpenAI has cried wolf on 'state-of-the-art' enough times that I want independent benchmarks before treating GPT-5.4 as anything more than a solid incremental upgrade dressed up as a leap.
Read more about this at: OpenAI