Gemini Deep Research API in 2026

You get rows, ready for your table

A Gemini report answers a question once, and someone still copies its facts into a table. Google's page says the Deep Research agent does not support structured outputs. Deep Search returns each match as a row with the columns you asked for: companies, people, filings, tenders, papers or products. get_find_decisions returns each condition's answer with its reason, the value, the source and the time it was read.

Source URL and read time on every value.

The price is set before the run

Google's figures are estimates: a typical Deep Research task uses about 80 searches, a Max task up to 160, and the agent decides how much to read. SpecStream quotes the most the whole job can cost, good for 30 minutes. compile reads your request into conditions and columns you can edit, quote_capacity_work prices it, and the run starts only when you approve that maximum as confirm_max_cents. The charge never goes past it.

You pay for rows, and the searches behind them are free

A row costs 5¢ with up to 5 facts, 15¢ with up to 15 facts and one set of related records, and $1 for a deep row with no limit. 10,000 rows with 5 facts cost $500. The searches and pages behind each row are not billed. Duplicates, rows you already own and rows a page contradicts are not charged.

Keep Gemini, and give it SpecStream's search

Gemini Deep Research takes remote MCP servers as tools, with headers for the key and allowed_tools to choose the tools. Add SpecStream at https://api.specstream.ai/mcp with search_web and read_pages, and the agent searches and reads the live web through SpecStream. A search costs $5 per 1,000 searches of up to 10 results, and a page costs $0.80 per 1,000 pages. The code is in the next section.

The Search API on the same key is the most accurate there is

The SpecStream Search API scores 97.9 F-score on all 1,000 official SimpleQA Verified questions (978 correct), and 64.0 item F1 and 43.4 row F1 on all 200 official WideSearch tasks. The Extract API scores 0.908 F1 on WCXB. Scores and method are on the benchmarks page. The Search API costs from $1 per 1,000 searches of up to 10 results, and the Extract API costs $0.80 per 1,000 pages.

The rows stay up to date and land in your tools

Save the request and it reruns hourly to yearly. Each row comes back labeled changed, unchanged or no longer matching. A changed or new row costs half its rate and an unchanged row costs 1¢, so rechecking 1,000 rows that did not change costs $10. New matches arrive in Slack, email, a sheet, a CRM or a signed webhook.

Perplexity Agent API

Sonar Chat Completions support ended on September 27, 2026, and Perplexity maps Sonar Deep Research to the Agent API's high effort level. You pay tokens at the model's rates plus tool fees: web_search costs $2.50 per 1,000 calls, or $1.00 on Fast Search. On SpecStream the searches and pages behind a row are not billed.

Parallel Task API

Priced per run by processor, from $5 per 1,000 runs on lite to $2,400 per 1,000 on ultra8x. The ultra processor costs $300 per 1,000, or 30¢ a run, and takes 5 to 25 minutes. SpecStream bills the rows a run delivers, and a run that finds nothing costs nothing.

Tavily Research

One request takes 4 to 110 credits on mini and 15 to 250 on pro, at $0.008 a credit pay as you go, so 3¢ to $2 a request. The charge lands between the minimum and the maximum. SpecStream names the most a run can cost before it starts.

OpenAI deep research

OpenAI shut down o3-deep-research and o4-mini-deep-research on July 23, 2026, and its deprecations page names gpt-5.6-sol as the replacement. Web search there costs $10.00 per 1,000 calls plus search content tokens at model rates. SpecStream takes the same research request in plain words and returns rows at 5¢, 15¢ or $1 each.