Deep research API benchmarks in 2026
Scores, who ran them, cost.
The deep research scores published in 2026 come from vendors' own runs, and each run puts its vendor first. On DeepSearchQA, Parallel's run scores its Ultra4x 86% and GPT-5.6 Sol 85%. On DRACO, Valyu's run scores Valyu 72.7% and Perplexity 70.5%.
SpecStream is the live web data engine for teams and agents. Its Search API scores 64.0 item F1 on all 200 official WideSearch tasks, above every system in the WideSearch paper. Deep Search returns research as rows at 5¢ to $1 each.
New accounts get $10 of credit, no card.
Every published deep research score, by benchmark
| SpecStream | Parallel | OpenAI | Gemini | Perplexity | Exa | Valyu | |
|---|---|---|---|---|---|---|---|
| The product scored | The Search API, with Deep Search and the Extract API on the same key | The Task API, processors from lite to ultra8x | GPT-5.6 Sol with web search. o3-deep-research shut down on July 23, 2026 | Gemini 3.1 Pro, and the Deep Research agent in preview | Perplexity's research at its high effort and its Deep Research | Exa Agent Max, and the deep-reasoning search type | DeepResearch, run on its Heavy mode |
| DeepSearchQA, 100 questions (Parallel's run, Aug 26, 2026) | Not in Parallel's run | 76% on Lite at $5 per 1,000 questions, up to 86% on Ultra4x at $1,200 | 85% on GPT-5.6 Sol PTC max, at $1,047.80 per 1,000 | 77% on Gemini 3.1 Pro high, at $123.90 per 1,000 | 68% on high, at $371.90 per 1,000 | 65% on Exa Agent Max, at $1,506.50 per 1,000 | Not in Parallel's run |
| BrowseComp, 100 questions (the same run) | Not in Parallel's run | 88% on Lite at $5 per 1,000 questions, up to 94% on Ultra4x at $1,200 | 85%, at $791.10 per 1,000 | 72%, at $194.40 per 1,000 | 86%, at $441.50 per 1,000 | 78%, at $1,043.50 per 1,000 | Not in Parallel's run |
| DRACO, 100 report tasks (Valyu's run) | Not in Valyu's run | 51.2% on Ultra8x, at $2,400 per 1,000 | 52.1% on deep research with o3, about $3,500 per 1,000 | 59.0% on Deep Research, about $4,500 per 1,000 | 70.5% on Deep Research with Opus 4.6, about $5,500 per 1,000 | 25.9% on Deep Reasoning, at $15 per 1,000 | 72.7% on Heavy, at $2,500 per 1,000 |
| WideSearch, 200 tasks, item F1 | 64.0 item F1 and 43.4 row F1, our run of Nov 8, 2026 | Scored as a search API on one shared agent, on the benchmarks page | 57.3 for o3 in a multi-agent setup, in the WideSearch paper | 57.4 for Gemini 2.5 Pro in a multi-agent setup, in the WideSearch paper | Scored as a search API on one shared agent, on the benchmarks page | Scored as a search API on one shared agent, on the benchmarks page | Not in the WideSearch paper |
| List price | 5¢, 15¢ or $1 a delivered row. Search API $5 per 1,000 searches, or $1 with speed=fast | $5 (lite) to $2,400 (ultra8x) per 1,000 runs | GPT-5.6 Sol: $4 input and $20 output per 1M tokens, plus $10 per 1,000 web searches | Google estimates $1.00 to $3.00 a Deep Research task | Tokens and tool calls. The cost varies with the model, the tokens and the tools used | deep $12 and deep-reasoning $15 per 1,000 requests | 8 credits (instant) to 2,500 credits (max) a task |
| Can you rerun it | Yes. The WideSearch tasks and scorer are public, and every raw answer and score of ours is on Hugging Face | Parallel's article gives a method to run on your own queries, and calls its figures a snapshot of that day | Scored in Parallel's and Valyu's runs | Scored in Parallel's and Valyu's runs | Scored in Parallel's and Valyu's runs | Scored in Parallel's and Valyu's runs | Valyu publishes its datasets, run scripts, raw outputs and judge evaluations on GitHub |
How to read a deep research benchmark
Every run puts the vendor that ran it first
Parallel's run of DeepSearchQA and BrowseComp puts the Parallel Task API first, and its article says so plainly: "You should know we're biased." Valyu's DRACO run puts Valyu first, on its Heavy mode. Linkup ranks its own /research endpoint first on SealQA-0 at 61%. Our WideSearch run is ours too, so every raw answer and score is public for you to check.
Read the cost per 1,000 tasks beside the score
The top score is rarely the one to buy. In Parallel's run, Task API Lite scores 76% on DeepSearchQA at $5 per 1,000 questions, and GPT-5.6 Sol scores 85% at $1,047.80: nine points for about 200 times the cost. In Valyu's run, Exa Deep Reasoning costs $15 per 1,000 tasks and scores 25.9% on DRACO, while Perplexity's Deep Research costs about $5,500 and scores 70.5%. Price the run you would buy, at the tier you would use.
Each benchmark measures a different job
DeepSearchQA asks multi-step questions. BrowseComp, created by OpenAI, tests persistent browsing for hard-to-find facts. DRACO, Perplexity's open benchmark, grades 100 long reports across 10 professional domains against expert rubrics. WideSearch, from ByteDance Seed, asks an agent to collect a whole table and scores every row and every cell. If your research ends in a table, WideSearch is the benchmark that measures it.
Our scores, and how to rerun them
On all 200 official WideSearch tasks the SpecStream Search API scores 64.0 item F1, 43.4 row F1 and a 7.0 success rate, the highest WideSearch score any search API has published. It scores 97.9 F-score on all 1,000 official SimpleQA Verified questions (978 correct), and the Extract API scores 0.908 F1 on WCXB. The tasks and scorer are on GitHub, our raw answers are on Hugging Face, and the method is on the benchmarks page.
Test your own job in minutes, for nothing
A benchmark tells you who did well on someone else's questions. Write 25 rows you already know the answers to as one request, send it to Deep Search, and read back how SpecStream reads it and the most it can cost, before you spend anything. Approve it and compare each row, with its source URL and read time, against your own answers. You pay 5¢, 15¢ or $1 for each row delivered, and a run that finds nothing costs nothing.
Other deep research APIs with published scores
Linkup
Linkup ranks its /research endpoint first on SealQA-0 at 61%, in its own run, at a fixed $0.25 to $2.50 a request. It returns a sourced answer. SpecStream returns every match as a row with a source on each value.
Tavily Research
Tavily Research on pro scores 38.8% on DRACO in Valyu's run, at about $1,060 per 1,000 tasks. It charges 4 to 250 credits a request at $0.008 a credit. SpecStream quotes the most a run can cost before it starts.
You.com Research
You.com Research on exhaustive scores 53.0% on DRACO in Valyu's run, at $450 per 1,000 tasks. It returns a report. SpecStream Deep Search returns structured rows at 5¢, 15¢ or $1 each.
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 run your own test on SpecStream
Sign up and you get $10 of credit, no card: up to 200 rows at 5¢, or 2,000 searches. Or start with no account at all. This call spends nothing: the 402 reply shows how SpecStream read the request, the columns and the most it can cost.
curl -i https://api.specstream.ai/v1/agent/lists \
-H "Content-Type: application/json" \
-d '{"request": "AI startups that raised a seed or Series A round in 2026, with the founders, the round size, the lead investor and the product in one line", "count": 25}'To test the Search API inside your own research agent, send each question as one call. The reply holds the answer, the page it was read from and every source.
curl "https://api.specstream.ai/v1/web-search?q=Who+led+the+seed+round+of+the+company+that+makes+Cursor" \
-H "Authorization: Bearer $SPECSTREAM_API_KEY"Over MCP at https://api.specstream.ai/mcp, your agent calls search_web for one fact, and compile, quote_capacity_work and authorize_capacity_work for a list, then reads the rows with list_find_rows.
Frequently asked questions
Something we didn't answer? Talk to a human
Which deep research API scores highest in 2026?
What is DeepSearchQA?
What is DRACO?
How much does a deep research run cost?
Has SpecStream been benchmarked?
Can I rerun a deep research benchmark myself?
Sources
SpecStream wrote this page. Each score of another company comes from the page of the company that ran the test, and each price from that company's own page.
- SpecStream benchmarksspecstream.ai/benchmarks, read Nov 12, 2026
- SpecStream pricingspecstream.ai/pricing, read Nov 12, 2026
- SpecStream for agentsspecstream.ai/agents.md, read Nov 12, 2026
- SpecStream raw benchmark answershuggingface.co/datasets/SpecStreamAI/benchmark-results, read Nov 12, 2026
- Parallel: Best deep research APIs in 2026, a benchmark reportparallel.ai/articles/best-deep-research-apis.md, read Nov 12, 2026
- Valyu: Introducing DeepResearchvalyu.ai/blog/introducing-deepresearch, read Nov 12, 2026
- Linkup: Best deep research API in 2026linkup.so/blog/best-deep-research-api-in-2026-openai-gemini-and-linkup-compared, read Nov 12, 2026
- WideSearch paperarxiv.org/html/2508.07999, read Nov 12, 2026
- WideSearch tasks and scorergithub.com/ByteDance-Seed/WideSearch, read Nov 12, 2026
- OpenAI o3-deep-researchdevelopers.openai.com/api/docs/models/o3-deep-research, read Nov 12, 2026
- OpenAI pricingdevelopers.openai.com/api/docs/pricing, read Nov 12, 2026
- Parallel pricingdocs.parallel.ai/getting-started/pricing.md, read Nov 12, 2026
- Gemini Deep Researchaistudio.google.com/docs/deep-research, read Nov 12, 2026
- Perplexity pricingdocs.perplexity.ai/guides/pricing, read Nov 12, 2026
- Exa pricingexa.ai/docs/admin/pricing.md, read Nov 12, 2026
- Valyu pricingdocs.valyu.ai/pricing.md, read Nov 12, 2026
- Tavily API creditsdocs.tavily.com/documentation/api-credits, read Nov 12, 2026