Service Status

A slow site no longer fails a whole /v1/fetch request with a 504: each URL gives up after 50s and comes back as success: false, refunded. Extract answers an unreachable or refused target page with 422 instead of 502, and invalid parameters (an unknown topic or search_depth, a malformed freshness) are now a 422 on every endpoint instead of being ignored or reported as an internal error. MCP validates arguments exactly like REST, a search with no answer to give no longer bills the include_answer add-on, research that finds no sources is refunded, refunded free credits go back to your free balance, and when search capacity is briefly unavailable you get a 503 to retry.

Searches with topic: "news" had been returning results with an empty description and no published_date, so a search_depth: "snippets" news search carried no text at all (at basic depth the page content still came through). Both fields are populated again for news results.

A query that matches nothing is retried once with its site: operator dropped (else its quotes), and the response sets relaxed_query to the query the results actually answer.

The credit price halved from $0.005 to $0.0025 and every credit cost doubled, so fetch, extract, research and the AI answer cost the same in dollars as before. Basic search dropped from 3 credits to 2, so a search is now $0.005 instead of $0.015, and snippets is $0.0025. Existing balances were doubled, and the free tier went from 1,000 to 2,000 credits.

GET /v1/status and GET /v1/updates now appear in skill.md and llms.txt so agents can discover them without scraping a page. Both are auth-free: poll /v1/status before running expensive workflows, poll /v1/updates to track changes to the API.