The best search API for AI agents
An agent acts on whatever its search returns, so accuracy is the number to choose on. SpecStream leads WideSearch (60.9, on the low-cost model; the best other API scores 57.6 even on the larger model) and SimpleQA Verified (99). See the benchmarks. That makes the SpecStream Search API the most accurate web search API in the world, at $5 per 1,000 searches. Each reply carries one answer with the page it came from and the time it was read, ranked rows, and the sources behind them. The same key reads pages with the Extract API and builds structured rows with Deep Search. SpecStream is the live web data engine for teams and agents.
New accounts get $10 of credit for their first 30 days, no card.
Six search APIs for agents, side by side
| SpecStream | Exa | Tavily | Parallel | Perplexity | Brave | |
|---|---|---|---|---|---|---|
| WideSearch, low-cost model | 60.9 | 53 | 47.9 | 45.5 (Fast), 44 (Turbo) | 47 | Not in the published scores |
| SimpleQA Verified, larger model | 99 | 91 | 92 | 97 (Advanced, Basic) | 95 | Not in the published scores |
| SimpleQA Verified, low-cost model | 96 | 91 | 94 | 94 (Fast), 91 (Turbo) | 94 | Not in the published scores |
| Price per 1,000 searches | $5 | $7 (auto, fast), $4 (lite, instant) | $8 (basic), $16 (advanced) | $5 (basic, advanced), $1 (turbo, fast) | $5, $1 (Fast Search) | $5 |
| Price per 1,000 pages read | $0.80 with the Extract API; a page with no text is not charged | $1 (Contents) | $1.60 (1 credit per 5 URLs) | $1 (Extract) | $0.50 (fetch_url in the Agent API) | Not stated |
| Work beyond one search | Deep Search: structured rows at 5¢, 15¢ or $1 a delivered row | Agent, $0.012 to $1.00 a request | Research, 4 to 250 credits a request | Task API, $5 to $300 per 1,000 runs | Agent API, no price per effort stated | Not stated |
| What failed or empty work costs | Nothing: a 503 is not charged, a page with no text is not charged, and a run that finds nothing costs nothing | Not stated | Not stated | Not stated | Not stated | Not stated |
| Remote MCP server | api.specstream.ai/mcp | mcp.exa.ai/mcp | mcp.tavily.com/mcp/ | search.parallel.ai/mcp (labeled legacy) | api.perplexity.ai/mcp | None stated; a local npx server |
Why agents run on SpecStream
The right answer more often
On WideSearch, an agent on the low-cost model scored 60.9 with SpecStream. The best other API in the published scores reached 57.6, and only on the larger model. On SimpleQA Verified, SpecStream scored 99 on the larger model and 96 on the low-cost one. Each wrong fact an agent takes from search turns into a wrong step, a retry, or a bad row, and that is where the cost of a cheap search shows up.
A reply the agent can act on
GET /v1/web-search returns an answer with its value, the page it came from and seen_at, then ranked rows and the sources behind them. When the sources do not agree, answer is null, and the agent knows to look further. Add passages=true and the reply also carries the text of the top pages, about a second later.
Source URL and read time on every value.
One key for search, pages, and rows
The Extract API reads 1 to 20 pages per call at $0.80 per 1,000, and a page with no text is not charged. Deep Search turns a request in plain words into structured rows at 5¢, 15¢ or $1 a row, plus 5¢ for an email and 15¢ for a phone. It shows the most a run can cost before it runs, reruns on a schedule from hourly to yearly, and delivers to the agent, a sheet, a CRM, Slack, or a signed webhook.
How SpecStream scores
The same numbers as the benchmarks page, from the official tasks and scorers.
| SpecStream |  Parallel |  Exa |  Tavily |  Perplexity | |
|---|---|---|---|---|---|
| Web researchWideSearch, GPT-5.6 Luna | 60.9 | 45.5 | 53.0 | 47.9 | 47.0 |
| Factual accuracySimpleQA Verified, GPT-5.6 Luna | 96 | 94 | 91 | 94 | 94 |
| Factual accuracySimpleQA Verified, GPT-5.6 Sol | 99 | 97 | 91 | 92 | 95 |
How to start
New accounts get $10 of credit for their first 30 days, no card. That is 2,000 searches. Create an API key and send the agent's first search.
Agents that speak MCP add https://api.specstream.ai/mcp and call the search_web tool. Guides for the OpenAI API, LangChain, LlamaIndex, CrewAI, and the Vercel AI SDK are in the framework docs.
curl -G https://api.specstream.ai/v1/web-search \
-H "Authorization: Bearer $SPECSTREAM_API_KEY" \
--data-urlencode "q=Who is the CEO of Stripe?"Frequently asked questions
Something we didn't answer? Talk to a human
What is the best search API for AI agents?
How much does the SpecStream Search API cost?
Is there a lower price per search?
Does SpecStream work over MCP?
https://api.specstream.ai/mcp over Streamable HTTP, with OAuth or a Bearer key. search_web runs the Search API, read_pages runs the Extract API, and the Deep Search tools quote, run, and read rows.What happens when a search fails?
Retry-After header and nothing is charged. The SDKs retry once when the wait is 10 seconds or less, and an Idempotency-Key header makes your own retries safe.Sources
SpecStream wrote this page. Each rival figure comes from that company's own page.
- SpecStream pricingspecstream.ai/pricing, read Nov 8, 2026
- SpecStream for agentsspecstream.ai/agents.md, read Nov 8, 2026
- Other providers' benchmark scores, Sep 9, 2026parallel.ai/benchmarks, read Nov 8, 2026
- Exa pricingexa.ai/pricing, read Nov 8, 2026
- Exa MCPdocs.exa.ai/reference/exa-mcp, read Nov 8, 2026
- Tavily pricingtavily.com/pricing, read Nov 8, 2026
- Tavily API creditsdocs.tavily.com/documentation/api-credits, read Nov 8, 2026
- Tavily MCPdocs.tavily.com/documentation/mcp, read Nov 8, 2026
- Parallel pricingdocs.parallel.ai/getting-started/pricing, read Nov 8, 2026
- Parallel Search MCPdocs.parallel.ai/integrations/mcp/search-mcp, read Nov 8, 2026
- Perplexity pricingdocs.perplexity.ai/getting-started/pricing, read Nov 8, 2026
- Perplexity MCP serverdocs.perplexity.ai/docs/getting-started/integrations/mcp-server, read Nov 8, 2026
- Brave Search APIbrave.com/search/api/, read Nov 8, 2026
- Brave Search MCP servergithub.com/brave/brave-search-mcp-server, read Nov 8, 2026