Hyperresearch is an AI deep research API and web console from Thalient Labs Inc., which operates from Kalispell, Montana. One question goes in and a reviewed report comes out, with each cited claim checked against the source it cites and every source kept in a searchable workspace vault. Founder Jordan Gibbs built it on his open-source hyperresearch command-line tool, which is MIT licensed and has 4.3k GitHub stars. The hosted service opened sign-ups on September 27, 2026. It is unrelated to HyperRESEARCH, the qualitative data analysis software published by ResearchWare, Inc.
A Deep run splits the question across 3 to 5 research chapters, reads wide, maps where the sources disagree and sends parallel investigators down each disagreement, then plans the report section by section and writes it in one pass. Before a report ships, each cited claim is read against its source, and a retracted paper the report does not acknowledge blocks it from shipping. Alongside the report a buyer gets the source list, a table of claims with the quote behind each, and on Deep a verification receipt recording which citations were checked and what the checker found. Each run searches the vault first, then the public web, logged out, and more than 40 data APIs such as SEC EDGAR, Census, BLS, OpenAlex and PubMed. Reports can be written in English, Chinese, Japanese or Korean.
Integration depth: Hyperresearch publishes a hosted remote MCP server at mcp.hyperresearch.ai/mcp. Claude, Cursor, ChatGPT and other MCP clients connect to it with browser OAuth and per-scope consent, so an agent can search the vault, read sources, start runs and check citations. The same runs, vault and reports are available over a REST API with an OpenAPI 3.1 document, scoped keys and webhooks. A standalone verification API checks citations in documents Hyperresearch did not write, returning a verdict per citation, quote integrity, a retraction sweep and an independence audit. Answers MCP is a separate server at answers.hyperresearch.ai/mcp with two tools: ask returns one short answer checked against word-for-word quotes, and search returns ranked sources. It is also packaged for LangChain and the Vercel AI SDK. There is no Google Drive or SharePoint connector, and a customer's own material reaches a run as uploads.
Pricing, in US dollars: runs are charged flat per run, with no subscription or seat licence. Light is $9 for one focused question, with about 20 to 50 sources and about 2,500 to 4,000 words in about 30 to 40 minutes. Deep is $49 for a broader question, with about 150 to 200 sources in 3 to 5 hours. Hyperresearch states that these figures are not guarantees. Adding a card makes the first Deep run free, once per account and once per card. Reruns cost 25%, 50% or 100% of the run price, depending on the step they restart from. Answers is billed per call from prepaid credit, with ask and search each from 5 cents, $5 of free credit on every account and top-ups from $10. Enterprise needs such as a HIPAA business associate agreement, single sign-on, a private document vault or retention terms are arranged by conversation rather than through a published plan.
What the entry price does not buy: a Light report's cited claims are checked, but the verification receipt comes with Deep. A new account cannot start runs until a card is on file. Scholarly lookups, vault fetches, verification pairs and vault storage are metered add-ons, with 500 scholarly lookups and 5 GB of storage included each month. A run takes tens of minutes to hours, so Hyperresearch states it is not a fit for quick lookups or live data, and it does not read pages behind a login. The terms state that reports are machine-generated and that verification is not designed to catch every error.
On data, Hyperresearch states that it does not train on customer content itself and does not license it to anyone for training. Questions and source text go to model providers through OpenRouter, and the policy names Meta's muse-spark contributor model for every pipeline role, Anthropic's Claude Opus 5.5 for planning, writing and final edits, and OpenAI as a fallback. Founder Jordan Gibbs confirmed to us in November 2026 that runs use Meta's contributor model today, that Meta may use what that step receives to train its models, and that the company is moving that step to a model that does not train on API data. Each provider's own terms apply, and a customer can ask for a specific provider to be excluded from its workspace. The privacy policy covers GDPR and CCPA rights, and the trust page states that SOC 2 readiness work is planned, with data processing agreements and residency handled on request.
Hyperresearch suits analysts, consultants and agent builders whose reports will be checked by someone else and who want the evidence kept for the next question. People who already pay for ChatGPT and want a cited report in well under an hour, drawing on connected apps, should look at ChatGPT Deep Research. Developers who need search results in seconds inside their own product should look at Exa AI. Researchers working mainly from peer-reviewed papers should look at Consensus.
Current state Q4 2026: in its own October 2026 test of 10 DRACO benchmark tasks with one judge, Hyperresearch reports that Deep scored 77.7 against 73.0 for Claude Research (Opus), and states that this is its own measurement and not an official leaderboard. In its own October 2026 testing it reports that Answers got 89 of 100 BrowseComp questions and 85 of 100 DeepSearchQA questions right, and it notes that Answers searches through Parallel's, Exa's and Valyu's APIs. Hyperresearch launched on Product Hunt in November 2026.
