Search API migration guides

One line changes: the address

Each route reads the request your client sends today and answers in the JSON it reads today. This is a Google Custom Search call, before and after.

# Before
curl "https://www.googleapis.com/customsearch/v1?key=$GOOGLE_API_KEY&cx=$CX&q=who+founded+Mistral+AI"

# After
curl "https://api.specstream.ai/customsearch/v1?key=$SPECSTREAM_API_KEY&cx=$CX&q=who+founded+Mistral+AI"

Every search on a route is one web search on your key, with the same limits and the same history as GET /v1/web-search. Parameters a route does not use, such as Google's cx, are accepted, so your code sends them as before.

What comes back, and what does not

The reply keeps the old API's field names: Google's items[] with title, link and snippet, Bing's webPages.value[], Tavily's and Exa's results[], Brave's web.results[], Serper's organic[], Firecrawl's data.web[] and SearXNG's results[]. Each reply also carries a specstream field your parser can ignore.

Three fields do not come back today: Tavily's published_date, Brave's page_age, and Exa's score, summary, image and favicon. Exa's author is null. Each guide lists every parameter and field of its API.

$0.005 a search, and a failed call costs nothing

A search of up to 10 results costs $0.005, which is $5 per 1,000. Each started block of 10 more results adds $0.006, up to 100 results a search. A call stopped by a 429 charges nothing, and credit you buy never expires.

A new account gets $10 of credit with no card, which covers 2,000 searches.

The most accurate web search API, at about one second a call

The SpecStream Search API scores 97.9 F-score on all 1,000 SimpleQA Verified questions, 978 of them correct, and 64.0 item F1 and 43.4 row F1 on all 200 WideSearch tasks. The Extract API scores 0.908 F1 on WCXB. The method and each score are on the benchmarks page.

Every drop-in call runs the fast search unless it asks for page text, the answer or results past the first ten. On November 9, 2026 production answered 718 of 720 searches at 1 to 20 a second, with a median of 0.82 s at 1 a second and 1.07 s at 20 a second, and the 95th percentile between 1.09 and 1.73 s. One account takes up to 8,800 searches a minute; a call over that gets a 429 with Retry-After and is not charged.

Nothing is kept on a zero retention key

Create a key with zero_retention and a search made with it keeps no request text and no answer after the reply, in the history or in a log line. spend_cap_cents is the most the key can spend.

curl -X POST https://api.specstream.ai/v1/account/keys \
  -H "Authorization: Bearer $SPECSTREAM_API_KEY" -H "Content-Type: application/json" \
  -d '{"name": "production", "scopes": ["account:read", "finds:spend"], "spend_cap_cents": 500000, "zero_retention": true}'

Pages and finished rows on the same key

The same key reads up to 20 pages or documents a call with the Extract API at $0.80 per 1,000 pages, and a page with no text is not charged. Deep Search returns every match for a request in plain words as structured rows, at 5¢, 15¢ or $1 a row, with a source URL on every value.

Vertex AI Search

Google's suggestion on the Custom Search page, for searching up to 50 domains. For full web search the page asks you to contact Google and express your interest. SpecStream searches the whole live web from the first call, on the path your code already uses, at $5 per 1,000 searches.

Grounding with Bing Search

Microsoft's path after the Bing Search APIs: a tool inside Azure AI agents at $14 per 1,000 transactions, and developers do not get the raw content it returns. SpecStream answers the Bing v7 call your code sends, with each result's name, URL and snippet, at $5 per 1,000.

Parallel Search API

Web search from $1 per 1,000 requests, with a guide to move Exa, Tavily and Brave calls to its own Search API, so the call and the parser are rewritten. On SpecStream the call and the JSON stay as they are, Parallel's own included: the Parallel migration guide changes one address. The score is 97.9 on all 1,000 SimpleQA Verified questions.