Skip to the page

The most accurate web search API in the world

SpecStream leads WideSearch and SimpleQA Verified, the standard benchmarks for search agents. On WideSearch, even a low-cost model on SpecStream beats every other search API running on a frontier model, at a fraction of the cost.

Why SpecStream leads

No other search API works the way ours does. SpecStream reads the live web at the moment you ask, goes to each site by its own road, and hands the agent the evidence with its source.

WideSearch

WideSearch, from ByteDance Seed, gives an AI agent broad research tasks and scores how much of the answer it collects correctly.

WideSearch
ProviderAgentScoreCost per 1,000 tasks
SpecStreamGPT-5.6 Luna60.9$39.40
Parallel AdvancedGPT-5.6 Sol57.6$692
Exa AutoGPT-5.6 Sol55.9$1,061
TavilyGPT-5.6 Sol55.9$1,072
Parallel BasicGPT-5.6 Sol55.3$965
PerplexityGPT-5.6 Sol53.5$547
Exa AutoGPT-5.6 Luna53.0$41.20
TavilyGPT-5.6 Luna47.9$107
PerplexityGPT-5.6 Luna47.0$24.80
Parallel FastGPT-5.6 Luna45.5$10.50
Parallel TurboGPT-5.6 Luna44.0$10.10

SimpleQA Verified

SimpleQA Verified asks short factual questions with one verified answer each.

SimpleQA Verified
ProviderAgentAccuracy
SpecStreamGPT-5.6 Sol99
Parallel AdvancedGPT-5.6 Sol97
Parallel BasicGPT-5.6 Sol97
SpecStreamGPT-5.6 Luna96
PerplexityGPT-5.6 Sol95
Parallel FastGPT-5.6 Luna94
PerplexityGPT-5.6 Luna94
TavilyGPT-5.6 Luna94
TavilyGPT-5.6 Sol92
Parallel TurboGPT-5.6 Luna91
Exa AutoGPT-5.6 Luna91
Exa AutoGPT-5.6 Sol91

Method

  1. WideSearch by ByteDance Seed. Item F1 on a 100-task sample, 50 in English and 50 in Chinese, on GPT-5.6 Luna at low reasoning, with the benchmark's own scorer. Cost covers model tokens and tool calls at API rates. Run November 5, 2026.
  2. SimpleQA Verified. Accuracy on 100 questions, on GPT-5.6 Luna at low reasoning and GPT-5.6 Sol at high reasoning. Run November 5, 2026.

Run it on your own work

Sign up for $10 of work, no card, and run your first request in the workspace, through the API or over MCP.