Baidu AI Search (Qianfan) compatible

The call, before and after

# Before
curl -X POST https://qianfan.baidubce.com/v2/ai_search/web_search \
  -H "Authorization: Bearer $BAIDU_API_KEY" -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Who founded Mistral AI and when?"}], "resource_type_filter": [{"type": "web", "top_k": 10}]}'

# After
curl -X POST https://api.specstream.ai/qianfan/v2/ai_search/web_search \
  -H "Authorization: Bearer $SPECSTREAM_API_KEY" -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "Who founded Mistral AI and when?"}], "resource_type_filter": [{"type": "web", "top_k": 10}]}'

The body and the parser stay. The host changes to api.specstream.ai/qianfan, and the key changes. The results come from the SpecStream Search API, in Baidu's JSON.

An AppBuilder client that sends its key as X-Appbuilder-Authorization: Bearer is read too.

Every parameter the route uses

Used in the search: messages (the search is the content of the last message), resource_type_filter (web with top_k 0 to 50, image with top_k 0 to 30), search_recency_filter (week, month, semiyear or year) and safe_search.

In search_filter: match.site keeps results to the sites you name, block_websites leaves sites out, and range.page_time keeps a window of days when you send both ends, with gte or gt and lte or lt. Each end is a day as YYYY-MM-DD or a preset such as now/d, now-1w/d, now-1M/d or now-1y/d.

One call is one search: the web results when it asks for any, and the image results when it asks for images alone. video and aladdin are accepted with their top_k and answered with no result. edition, search_source and query_policy are accepted, and the price is the same for every edition.

Refused with HTTP 400 and Baidu's InvalidArgument: a top_k above the kind's maximum, a resource type other than web, image, video or aladdin, a search_recency_filter Baidu does not take, and a call that asks for neither web nor image results.

Every field that comes back

The reply has request_id and references. Each reference has id (counted from 1), title, url, web_anchor, website, snippet, content, date as YYYY-MM-DD HH:MM:SS when the page gives a time, type (web or image) and rerank_score.

content carries the result's snippet text. On a web reference icon, image and video are null. An image reference has image with url, height and width.

One field is added: specstream, with what the Search API returns beyond Baidu's fields, including the call's price_usd. A parser that reads Baidu's fields never sees it.

A refusal carries Baidu's requestId, code and message and keeps its HTTP status, so a client's status check sees it: 216003 with 401 for a key that cannot be read, BILLING_INSUFFICIENT_BALANCE with 402 and RATE_LIMIT_SEARCH_QPS with 429.

$1 per 1,000 searches, with no QPS cap

A call of up to 10 results costs $0.001. Each started block of 10 more results adds $0.0012: top_k 20 costs $0.0022 and top_k 50 costs $0.0058. A call with no resource_type_filter returns 20 web results, as on Baidu, so send top_k 10 to pay $0.001.

The site filter and the date filter cost nothing extra. A call stopped by a 429 charges nothing, and credit you buy never expires.

The most accurate web search API, in about one second

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

Every call on this route runs the fast search. Measured on production on November 9, 2026, its median was 0.82 s at 1 search a second and 1.07 s at 20 a second, the 95th percentile stayed between 1.09 and 1.73 s, and it answered 718 of 720. It has no limit per key and no daily limit.

You decide what is kept, on each key

Create the key with zero_retention and a search made with it keeps no request text and no answer after the reply: it is not in the history, its answer is not reused for another call, and its log lines carry neither. 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}'

Search, pages and rows on one 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. The most a run can cost is approved before it starts.

Alibaba Cloud IQS Web Search

Alibaba Cloud's search API for agents, priced by engine: 8 CNY per 1,000 calls on China Edition LiteBasic, 42 CNY on the Generic edition's first tier at 30 QPS, and 56 CNY on Global Edition GlobalAdvanced. SpecStream answers its POST /search/unified as well, at https://api.specstream.ai/aliyun-iqs, for $0.001 a search of up to 10 results.

Tavily

A search API sold in credits at $0.008 each pay as you go: a basic search takes 1 credit and an advanced search 2. SpecStream answers Tavily's POST /search at https://api.specstream.ai/tavily for $0.001 a basic search and $0.005 an advanced one.

Brave Search API

A search API on Brave's own index at $5 per 1,000 requests, with $5 in free credits every month and 50 queries per second. SpecStream answers Brave's call at https://api.specstream.ai/res/v1/web/search at $1 per 1,000, a fifth of Brave's rate, and has no limit per key.