OpenAI released GPT-5.4 on March 5, in three versions: the standard model, GPT-5.4 Thinking for reasoning-heavy work, and GPT-5.4 Pro for maximum performance. The headline number is the context window — up to one million tokens in the API, the largest OpenAI has offered. Two other changes may matter more to people building products on it.
What Was Announced
- A 1M-token context window in the API, the largest OpenAI has offered so far.
- Tool Search. Instead of putting the definition of every available tool into each request, the model can look tools up when it needs them. OpenAI says this makes requests faster and cheaper in systems with many tools.
- Token efficiency. OpenAI says GPT-5.4 solves the same problems with significantly fewer tokens than its predecessor.
- Benchmarks. OpenAI reports a record 83% on its GDPval test of knowledge work, record scores on computer-use benchmarks, and individual claims 33% less likely to contain errors than GPT-5.2. These are vendor figures, and worth testing against your own tasks.
A Million Tokens Is Not a Retrieval Strategy
A million tokens is a lot — a large codebase, or several long contracts, in one request. It will make some designs simpler: fewer chunking pipelines, fewer cases where the answer sits in a part of the document you did not retrieve.
It does not make context free. Every token sent is paid for and adds latency, and long-context accuracy is something to measure on your own documents rather than assume. For most applications, selecting the right material still beats sending everything. The bigger window is most valuable as headroom: the rare request that genuinely needs it no longer fails.
Tool Search Is the Quiet Win
Agent systems tend to accumulate tools. Each tool needs a name, a description and a parameter schema, and all of that is sent with every call whether it is used or not. With dozens of tools, the definitions alone can rival the actual conversation in size.
Letting the model fetch tool definitions on demand addresses that directly. It also changes how tools should be written: if the model finds tools by searching, clear names and descriptions become even more important than before. A vague tool description used to cost a little accuracy; now it can make the tool hard to find at all.
What We'd Do This Month
- Re-run your own evaluation set before switching models. Fewer tokens per task and higher benchmark scores only matter if they hold on your prompts.
- Check whether long-context requests could replace a fragile retrieval step — and measure cost and latency before committing to it.
- If you run an agent with many tools, try Tool Search and tidy up tool names and descriptions while you are at it.