Modes
Four depths of search. Pick the cheapest one that answers your question.
Every search endpoint takes a mode. It sets how many candidates are judged and whether the best ones are
opened and read before they are ranked.
ultra | fast | normal | deep | |
|---|---|---|---|---|
| What each headline is judged on | headline | headline and standfirst | headline and standfirst | headline and standfirst |
| New headlines judged per search | up to 160 | up to 160 | up to 400 | all of them |
| Articles opened and read | — | — | up to 4 | up to 8 |
highlights | not available | not available | optional | on by default |
essential | optional | optional | optional | on by default |
| Typical latency | about 1 s | about 1 s | 2–5 s | 10–15 s |
| Price | Search | Search | Search | Deep search |
normal is the default.
Choosing a mode
ultrais the cheapest. Use it for high-volume classification, alerts, and anything where the headline says it all.fastis the best default for agents in a loop: quick, cheap, and good recall when the headline names the topic.normalreads the top results before ranking them. Use it when a wrong result is expensive — a headline that promises more than its article is caught and moved torejected.deepjudges the whole index and reads more. Use it for research, reports and questions where you need everything, with highlights and essentials included.
Results arrive in two stages
In every mode, results judged from the headline arrive first — about a second in — and each one is
confirmed in place as its article is read. With streaming you get those
partial results as partial events; without it, you get the final result.
What doesn’t cost twice
- Judgment memory. What the model already said about a headline for a query is reused for a week, so a repeated or similar search only pays for what is new.
- Result cache. An identical request within 10 minutes is served from cache:
usage.tokensis0and it is not billed.
See Costs and caching for hard caps per request.