Claude Research or Hyperresearch: which one fits the job

Use Claude Research when you already pay for Claude and want a cited answer in minutes, especially one that draws on your Gmail, Calendar or Google Docs. Use the Claude API with web search when you want to build your own research loop inside a product. Use Hyperresearch when the report itself is the deliverable: $9 or $49 flat, with its cited claims checked against their sources before it ships.

Claude Research is the strongest of the other research agents we have graded. On our ten DRACO tasks it scored 73.0 to Hyperresearch’s 77.7, second on the board and more than 11 points ahead of every other system, and it scored higher than us on five of the ten tasks. This page sets out where each one fits, using Anthropic’s own pages as of October 2026.

Claude Research, the Claude API and Hyperresearch side by side

Anthropic sells research two ways. Research is a mode in the Claude app on paid plans. The Claude API has no Research product, but it sells the parts: a web search tool, a web fetch tool and models you can run in your own loop. Hyperresearch sells one thing, a research run, through a REST API, an MCP server and a web console.

Anthropic’s cells come from Use research on Claude, When should I use web search, extended thinking, and research?, the Claude pricing page, the web search tool and web fetch tool docs, and custom connectors, all read on October 12, 2026. Hyperresearch’s source counts, lengths and times are the targets the pipeline gates on, not guarantees.

Claude Research costs nothing extra, but it spends your plan’s limits

Research is included in every paid Claude plan and missing from the free one. Anthropic’s pricing page lists Pro at $20 a month, or $17 a month billed annually at $200 up front; Max from $100 a month for 5x or 20x Pro’s usage; Team at $25 a seat a month ($20 billed annually) for a Standard seat and $125 ($100) for a Premium seat; and Enterprise at $20 a seat a month, billed annually, plus usage at API rates.

There is no per-report count. The Research help article says: “Research is subject to the same limits as standard Claude conversations. However, research sessions can use up your limits faster due to Claude retrieving multiple sources and providing comprehensive responses.” The pricing page explains those limits: they “reset on a rolling five-hour session window, and paid plans add weekly limits on top”, your chats, desktop, mobile and Claude Code all draw from the same pool, and “there’s no fixed message count”. Anthropic also reserves the right to add weekly and monthly caps or limit particular features “at our discretion”.

In practice, a heavy research afternoon spends the same budget as your coding and your everyday chats. When you hit the wall, Anthropic offers three ways out: wait for the reset, move to a higher plan, or, on paid plans, turn on usage credits billed at standard API rates. Deep research limit reached compares the same wall across ChatGPT, Gemini and Claude.

A Hyperresearch run costs $9 or $49 and draws on no plan allowance. The price comes back in the same response that creates the run, before any work starts, and runs that fail on our side are refunded automatically. If you run a few Research sessions a week inside a Pro plan you already have, Claude is far cheaper per report. Hyperresearch costs more per report and buys a larger source base, a claim check before release and a vault that keeps every source.

Claude Research has no API, but the Claude API sells the parts

Anthropic’s developer docs list no Research endpoint and no research agent as of October 2026. If you want Claude to research inside your own product, you build the loop from the parts Anthropic does sell:

• Web search tool. “$10 per 1,000 searches, plus standard token costs for search-generated content”, per the web search docs. Citations are always on. max_uses caps the searches in one request, and the docs say simple questions use 1 to 3 searches while “comparative or multientity research can use 10 or more”. Web search is not available on Amazon Bedrock.
• Web fetch tool. No charge beyond tokens, per the web fetch docs, which also note that it “does not support websites dynamically rendered with JavaScript”.
• Models. On the pricing page, Opus 5.5 costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens.
• Managed Agents. A hosted agent harness, in beta, “best for long-running tasks and asynchronous work”, at $0.08 per session-hour of active runtime plus tokens, per the overview.

The cost depends on what your loop does. As an illustration, a loop on Opus 5.5 that makes 50 searches, reads 1 million input tokens and writes 50,000 output tokens costs:

50 searches      x $10.00 / 1,000  = $0.50
1,000,000 in     x $4.00 / 1M      = $4.00
50,000 out       x $20.00 / 1M     = $1.00
                                     -----
                                     $5.50

Those counts are assumptions, not measurements. Cached input lowers the input line if your loop re-reads the same context. What you give up is the research behaviour itself: how the question is broken down, when to stop searching, and whether each citation supports its sentence are all yours to build. Moving off OpenAI’s deep research API walks through the same trade with OpenAI’s models.

The Hyperresearch call is one POST:

curl -X POST https://api.hyperresearch.ai/v1/runs \
  -H "Authorization: Bearer <YOUR_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "<question>",
    "tier": "premier",
    "webhook_url": "https://example.com/hooks/hyperresearch"
  }'

query takes up to 4,000 characters. tier is light or premier, which is Deep on the wire. The 202 response carries the run id and price_cents before any work starts. Every run is asynchronous: poll GET /v1/runs/{id} or wait for the webhook, then read the result for the report, the full source list, every source tried and skipped with a reason, the claims table and, on a Deep run, the verification receipt.

Claude Research accuracy for client work

Claude Research cites its sources, and Anthropic says its citations are “easy-to-check”. That is the right description: checking them is your job. We found no per-citation check result in Anthropic’s Research help pages, and no published citation accuracy rate for Research.

For client work, what matters is knowing which citations failed. A cited report that is right nine times in ten still needs someone to find the tenth before the client does.

A Hyperresearch run does part of that work before the report ships. Every cited figure is read against the source it cites, and what does not hold is corrected in one guarded revision. A Deep report adds a verification receipt with a verdict for each checked citation (supported, partially supported, unsupported or wrong source) and the passage from the source that the verdict rests on. A cited paper that has been retracted, or a quotation that does not appear in its source, stops the report from shipping.

A receipt is a record of checks, not a promise that every claim is true. A supported verdict means the source says what the sentence says, not that the source is right. Hyperresearch has not published a measured error rate for its own checker. A Light report gets the claim check but no receipt.

On ten DRACO tasks, Hyperresearch scored 77.7 and Claude Research 73.0

We ran Hyperresearch Deep, Claude Research and six other research agents on the same 10 tasks from DRACO, a deep research benchmark from Perplexity and Harvard, in October 2026. All were graded with the same judge, GPT-5.2 at the paper’s settings. Our scores and Claude’s are each the mean of two grades. Claude was run in the Claude app with no custom instructions, on Opus.

This is a close result, and Claude earned it. Claude scored higher than Hyperresearch on five of the ten tasks: Law, 96.9 against 90.6; UX design, 80.6 against 74.2; General knowledge, 69.0 against 65.9; and Academic and Shopping by under a point each. It also scored higher on the extra medicine task we ran outside the mean, a patient deciding whether to go to the emergency room, 90.2 against 89.5.

Most of the gap comes from one task. On the Medicine task, about monitoring, insurance authorization and pacing thresholds for a patient with a worsening heart block, Hyperresearch scored 72.2 and Claude 37.1. Hyperresearch was also ahead on Needle in a haystack, 77.6 against 64.2, and Personalized assistant, 64.1 against 54.6, and by smaller margins on Finance and Technology.

These are 10 of DRACO’s 100 tasks, two grades of the same report can differ by up to about 9 points, and this is our measurement, not an official leaderboard. A 4.7-point gap on ten tasks does not settle which tool writes the better report on your question. How we tested has every task score and the other systems we graded, or see a full Deep report.

Where Claude Research is the better choice

• Speed. Minutes against half an hour or more. If someone is waiting on the answer, Claude wins.
• Price, if you already pay for Claude. Research is part of Pro at $20 a month. A single $9 Light run costs nearly half of that.
• Your Google Workspace. Research can work across Gmail, Google Calendar and Google Docs when they are connected, alongside the web. Hyperresearch reads public pages logged out and has no Gmail or Drive connector; your own documents reach it only as uploads.
• Conversation. Research lives inside a chat, so you can push back, ask a follow-up and change direction in the same thread. A Hyperresearch run takes one question and returns one report.
• Quality on many tasks. On our DRACO tasks Claude scored higher than us on Law, UX design, General knowledge, Academic and Shopping. It is a strong research tool, not a cheap substitute.
• Building your own loop. If you want full control over how research is done, the Claude API gives you search, fetch and the models at published per-unit prices.

Where Hyperresearch is the better choice

Hyperresearch fits when the report is the thing you hand over and someone will ask where each number came from. A Deep run splits the question into 3 to 5 research chapters, reads about 150 to 200 sources, maps where they contradict each other, sends investigators down the disputed points, and plans the report section by section before one writer writes it in one pass from all of the evidence, followed by the claim check and receipt. How it works lists every stage.

It also fits when you do not want research to spend your chat budget. A Deep run is $49 whether it reads 150 sources or 200, and it never competes with your coding sessions for the same five-hour window.

The third reason is what stays behind. Every source a run reads is kept in your workspace vault with its provenance, and the next run in the same project searches the vault before the web. Claude’s Projects keep conversations, files and instructions. Hyperresearch keeps the source corpus itself, full-text searchable, and exports it as plain Markdown.

The fourth is a price per report through an API. Claude Research has none. Hyperresearch quotes the price before the run starts and refunds runs that fail on our side.

Using Hyperresearch from inside Claude

You do not have to choose one. Anthropic’s connectors article says custom connectors using remote MCP are available on Free, Pro, Max, Team and Enterprise plans, with Free limited to one. Add the Hyperresearch MCP server at https://mcp.hyperresearch.ai/mcp and Claude can search your vault, read the evidence behind a report, and start a run on your approval. Connect Hyperresearch to Claude has the steps for claude.ai, Claude Desktop and Claude Code.

A Deep run takes hours, longer than any MCP client will hold a tool call open, so the server starts a run, returns its price and an id, and lets Claude check back. Long-running MCP tools explains the pattern. Use Claude Research for the quick question in the chat and hand the client-ready report to a run.

If you would rather run everything locally, the open-source hyperresearch CLI runs the same pipeline inside Claude Code, on your own plan. Claude Code as a research harness sets it up.

How this comparison was made

Every Anthropic figure on this page comes from Anthropic’s own help center, pricing page and developer docs, read logged out on October 12, 2026 and linked where it is used. Where Anthropic does not publish a figure, such as sources per report or report length, the table says so instead of estimating. The API cost example is arithmetic on Anthropic’s published rates with assumed counts. The quality numbers are our own grading of both products on 10 DRACO tasks in October 2026, described on how we tested. Hyperresearch figures are the targets the pipeline gates on, as printed on pricing. For the same comparison against other tools, see Perplexity and Gemini. Anthropic changes its plans and limits often, so we will re-check these numbers by the end of 2026 and change the date at the top when we do.