Tool search: your tools no longer live in the prompt

On 5 March 2026, OpenAI stopped asking the prompt to carry the tool catalogue. There are not fewer tools: the place where they are declared has changed hands, and so has what you pay on every call.

What has shipped

gpt-5.4 and gpt-5.4-pro, on 5 March 2026. OpenAI presents them as models for professional work, with 1M context, built-in computer use, native compaction and tool search.

The three are not the same piece of news. Computer use brings interaction with an interface —screenshots, clicks— into the API, as the computer tool in the Responses API. Compaction turns a long agent prompt into a maintenance problem —what to keep— rather than a fitting problem. And tool search defers the tool catalogue to the moment of the call.

17 March brought gpt-5.4-mini and gpt-5.4-nano. The mini ships tool search and computer use; the nano ships compaction but no tool search.

What changes for reasoning

OpenAI gives three reasons for tool search: spend fewer tokens, preserve cache performance and answer faster. All three come out of the same move. Eighty definitions in every request are eighty definitions you pay for, that occupy the prefix and that have to be read before anything is decided.

The real change is where the index lives. A tool description used to be prompt text: you wrote it once and pasted it into every call, in the same file as the rest of the prompt. Now it is an artefact queried at runtime. You notice it on next month's bill. The list can no longer go stale inside the prompt, because it is not inside it. And the stable prefix stops dragging definitions that call is not going to use, which is what OpenAI calls preserving cache performance.

What does not change is the description. One that does not tell search_orders from lookup_order still picks wrong, whoever reads it. The work does not disappear: it moves from listing to describing.

What changes for writing prompts

A prompt that listed its tools breaks in three places, and none of them raises an error.

The list falls behind: a tool that is no longer in the catalogue stays in the text, and the model asks for it. The positional reference: "the third tool" was an address while you set the order, and now the search sets it. And the schema duplicated in prose: if you describe the parameters in the prompt and also in the definition, there are two versions, and the one that arrives is the definition.

What you write instead is the policy: when to search, with which words, what to do when nothing comes back and when to stop.

Tools: the catalogue is queried at runtime and does not go in this prompt.

1. Search with the verb and the object of the request —"search order",
   "refund order"—, not with the user's whole sentence.
2. Use the tool the search returns. Do not bend a similar one so that it
   fits.
3. If the search returns nothing usable, say what is missing and stop. Do
   not chain the closest match.
4. Before the result, one line: which tool you used and why.

And one rule that survives both regimes: ban inventing tools, not using one that is not on the list. The list is no longer the source of truth; the catalogue is.

What to stop doing

• Pasting the catalogue into the system block. That is what tool search exists to remove, and the fixed cost of every call goes with it.
• Duplicating the schema in prose. When the prompt and the definition disagree, the definition wins.
• Referring to tools by position. "The third one" only worked while the order was yours.
• Budgeting the prompt as if the definitions were part of it. With the catalogue out, the budget is counted on what is left.
• Assuming the whole family ships it. gpt-5.4-nano has no tool search: if its prompt relies on one, the catalogue comes back in through the side door.

Sources

• OpenAI API changelog
• Introducing GPT-5.4
• Introducing GPT-5.4 mini and nano