Guides
Deep research APIs with source-level citations, compared
For source-level citations with usage-based pricing, the best-documented choices on 2026-11-03 are Parallel’s Task API, which attaches citations with excerpts, reasoning and a confidence rating to each output field for a fixed price per run, and Valyu DeepResearch, which can attach the exact supporting passage to each numbered citation for a fixed credit price per task. Hyperresearch is the pick when the deliverable is a long report whose citations must be checked against their sources: a Deep run is $49 flat, and the Answers API returns a short answer with word-for-word quotes for 2¢ to 50¢ a call. OpenAI’s dedicated deep research models were shut down on July 23, 2026, and Perplexity’s Sonar Deep Research is now the high effort level of its Agent API.
We make Hyperresearch, so weigh that. Every vendor price, limit and model name here comes from that vendor’s own docs, read on 2026-11-03 and linked inline.
Seven APIs side by side
“Citation unit” is what one citation is attached to. “Excerpt” means the API returns the source text that supports the claim, not only a URL.
| API | Citation unit | Excerpt or check | Pricing model | Price, read 2026-11-03 | Typical time | Async |
|---|---|---|---|---|---|---|
| OpenAI Responses API + web search | Character span in the answer (url_citation) |
URL and title, no excerpt | Per token plus per search call | Model tokens plus $10 per 1,000 search calls | Several minutes | Background mode, webhooks |
Perplexity Agent API, high effort |
Inline citation on source-backed claims | No documented excerpt | Per token plus per tool call | Model tokens plus $2.50 per 1,000 standard searches | Not published | Background mode, poll or stream |
| Parallel Task API | Each output field, and each list element | Excerpts, reasoning, confidence | Fixed per run | $0.10 (pro) to $2.40 (ultra8x) per run |
2 min to 2 hr | Webhooks, task groups |
| Exa Agent | Text or structured fields, “when emitted” | Grounding object | Fixed per run, or metered | $0.012 to $1.00 per fixed-effort run; metered up to a $5 or $20 default cap | Async, varies | Poll or stream |
| Gemini Deep Research agent (preview) | Citations in the response | Not documented on the page | Per token plus tool fees | Google estimates about $1 to $3 a task, Max $3 to $7 | Most under 20 min, cap 60 min | Background, poll or stream |
| Valyu DeepResearch | Numbered marker in the report | Optional exact passage and page per source | Fixed credits per task | 100 credits (standard) to 2,500 (max) |
Limits 2 to 90 min by mode | Webhooks |
| Hyperresearch Deep run | Each citation in the report | Checked against its source, verdict per checked citation | Flat per run | $49 (Light $9) | 3 to 5 hours (target) | Webhooks |
Hyperresearch Answers ask |
The short answer | Word-for-word quote, checked against the page | Per call | 2¢, 10¢ (20¢ if it digs deeper), 50¢ | About 2 s to 5 min | Async with Prefer: respond-async |
OpenAI: deep research is now a mode, not a model
OpenAI no longer sells a deep research model. The deprecations page says access to o3-deep-research and o4-mini-deep-research “was shut down on July 23, 2026”, with gpt-5.6-sol listed as the replacement. The web search guide now describes deep research as a general reasoning model run at high reasoning effort with the web search tool, in background mode, and says it “can run for several minutes”.
Citations come back as url_citation annotations: a URL, a title and the character range of the answer they support (start_index, end_index). A separate sources field lists every URL the model consulted, which is usually more than it cited. There is no excerpt from the source, so checking a claim means fetching the page yourself.
Pricing is usage-based. On the pricing page, gpt-5.6-sol is $4.00 per million input tokens and $20.00 per million output tokens, at promotional rates “available at least through November 21, 2026”, and web search is $10.00 per 1,000 calls plus the search content tokens at model rates. A background response can notify you with a response.completed webhook. If you are moving off the old models, migrating off the OpenAI deep research API maps the calls.
Perplexity: Sonar Deep Research became the high effort level
Perplexity ended Sonar Chat Completions support on September 27, 2026. Synchronous calls are being reformulated as Agent API requests; asynchronous Sonar requests are no longer supported and must move to background mode. Sonar Deep Research maps to the high effort level, which Perplexity describes as exhaustive source coverage “with inline citations for source-backed claims”. It runs up to 40 steps with web search and URL fetch tools.
The Agent API is billed per token at the chosen model’s rates plus tool fees: $0.0025 per standard web search, $0.001 per fast search, $0.0005 per URL fetch. A completed response reports its own cost in usage.cost.total_cost. Long jobs run in background mode and you poll or stream them; webhook callbacks are on Perplexity’s roadmap, not shipped. Perplexity deep research compared goes further.
Parallel Task API: citations for every field
Parallel’s Task API has the finest-grained citations of the group. Each result carries a basis: for every output field, and for every element of a list field, a set of citations (URL plus excerpts), a reasoning note and a confidence of high, medium or low. Ask for a text schema instead and you get a Markdown report with in-line citations, with one basis entry for the whole output.
Pricing is per run, not per token or per field, and failed runs are not billed. Deep research sits in the pro and ultra processors: pro is $100 per 1,000 runs (2 to 10 minutes), ultra $300 (5 to 25 minutes), up to ultra8x at $2,400 (5 minutes to 2 hours). Completions arrive by webhook with HMAC signatures, and the default rate limit is 2,000 task runs a minute. The Parallel Task API guide covers task groups and schemas, and Parallel Task API vs Hyperresearch compares the two directly.
Exa: the Agent API replaced /research
Exa’s changelog records that the /research endpoint was retired, and points long-running research to Exa Agent. An Agent run returns a text answer, optional schema-checked JSON, and output.grounding, citations for text or structured fields “when emitted”. Runs are async: you stream events or poll.
Exa’s pricing page offers a fixed price per run by effort, from $0.012 (minimal) to $1.00 (xhigh), or metered runs at $0.10 per agent compute unit plus $0.005 per search, capped by default at $5 (auto) or $20 (ultra). For a quick cited answer, /answer is $5 per 1,000 requests. Exa alternatives compares Exa with other search APIs.
Gemini Deep Research agent: preview, billed by tokens
Google’s Deep Research agent is in preview and runs only through the Interactions API, as deep-research-preview-04-2026 or deep-research-max-preview-04-2026. You start it with background=true and poll or stream. Research is capped at 60 minutes and Google says most tasks finish within 20. Structured outputs are not supported. The docs tell you to review “the citations provided in the response” but do not document an excerpt per claim.
The agent is billed on its token use at standard Gemini rates plus tool fees. Google’s own estimates are about $1 to $3 a task for the standard agent and $3 to $7 for Max, which it calls “subject to change”. Gemini Deep Research compared has more.
Valyu DeepResearch: fixed credits per task, passages per citation
Valyu’s DeepResearch report cites its sources with numbered markers, and the sources list gives each title and URL. A source can also carry locations: the exact passage the citation rests on, with a page number when the document has pages. The released SDKs do not expose it yet.
Each task has a fixed price in credits that includes its searches and page reads: instant 8, fast 25, standard 100, heavy 500, max 2,500. Pay as you go is 1,600 credits per $10, so by our arithmetic a standard task is about $0.63 and a max task about $15.63; monthly plans lower the rate. Hard time limits run from 2 minutes (instant) to 90 minutes (max), and a webhook_url notifies you on completion. Valyu DeepResearch vs Hyperresearch compares the two.
Hyperresearch: checked reports and quoted answers
Hyperresearch has two APIs for cited research.
A Deep run (POST https://api.hyperresearch.ai/v1/runs, tier: "premier") takes a question of up to 20,000 characters and returns an adversarially reviewed Markdown report. Every citation is checked against the source it cites before the report ships, and the verification receipt records a verdict for each checked citation: supported, partially supported, unsupported or wrong source. A Deep run targets about 150 to 200 sources and 3 to 5 hours. It costs $49 flat, Light runs cost $9, and your first Light report is free when you sign up, with no card. Pass webhook_url to be called when it finishes. Every source a run reads is kept in a searchable vault.
The Answers API is for an agent that needs one fact, not a report. ask (POST /v1/answers) returns a short answer with the quotes it rests on, each checked against its page, or an honest not_found. It costs 2¢ (fast, usually a few seconds), 10¢ (standard, 20¢ if it digs deeper) or 50¢ (max, up to about 5 minutes), and auto picks one of the three and bills the mode that ran. fast answers from the open web only and does not query structured data sources; for a figure from a company filing, a government statistic or a registry (SEC EDGAR, FRED, Census, FDA, FEC, USPTO, ClinicalTrials.gov), use standard or auto. The sources page lists every structured source and which modes reach it. search returns up to 25 ranked sources with matching passages marked verified when cut word for word from the live page: 5¢ in standard (usually 5 to 10 seconds) or 15¢ in max (20 to 45 seconds). Both share a limit of 60 calls a minute per API key and 120 per workspace, and each account starts with $5 of credit.
On DeepSearchQA, Hyperresearch Search scores 89% and Exa scores 72%, with the same agent and the same grader. Hyperresearch Search was tested in standard mode as it runs in production.
Parallel Task API vs Perplexity deep research
These two get compared most, and they differ in what a citation is attached to. Parallel cites per field, with excerpts and a confidence level, so a pipeline can route a low-confidence field to review without reading the whole answer. Perplexity’s high effort cites inline in prose. Parallel charges a known price per run before you send it; Perplexity charges for the tokens and tool calls the run used and reports the cost afterwards. Parallel has webhooks and batch task groups; Perplexity has background mode with polling or streaming. For enrichment of many rows, Parallel fits. For a cited prose answer in a product that already uses Perplexity’s models and search, Perplexity fits.
How to choose
The criteria that separate the options:
- What you need to check. If a person or a pipeline has to confirm each claim, you need the supporting text, not only a URL. Parallel (excerpts per field), Valyu (optional passage per source), Hyperresearch (a checked verdict per citation, or a checked quote per answer) return it. OpenAI and Perplexity return the link and leave the check to you.
- Price you can quote before the run. Parallel, Valyu, Exa’s fixed efforts and Hyperresearch charge a fixed price per run, task or call. OpenAI, Perplexity, Gemini and Exa’s metered mode bill by usage, which tracks effort but is only known after the run.
- Output shape. For your own JSON schema, filled field by field, use Parallel or Exa Agent. For a long report, use Valyu, Gemini, Parallel’s text mode or Hyperresearch.
- Time. A caller waiting on a reply needs seconds: Answers
search, Exa/answeror Valyuinstant. A report can take minutes (Gemini, Parallelpro) or hours (a Hyperresearch Deep run). - Volume. For thousands of rows a day, Parallel’s task groups at 2,000 runs a minute are built for it. A Hyperresearch Deep run is not.
Where each other option is the better pick:
- Parallel for structured enrichment at volume, with a citation and a confidence level on every field.
- Valyu for a fixed-price report that draws on its paid datasets (filings, trials, patents on higher plans), with a passage behind each citation.
- Exa Agent for list building with a hard cost cap per run.
- Gemini for a fast, inexpensive report if you already build on Gemini and can accept a preview API.
- OpenAI for full control of the research loop inside an app that already runs on the Responses API.
- Perplexity for cited prose answers billed by usage, with one bill for model and search.
- Hyperresearch when the report goes to a client or a decision and every citation must be checked against its source first, or when an agent needs one quoted, checked fact per call.
The deep research pricing comparison covers the consumer plans as well, and cost per verified citation gives one unit for comparing these APIs on price, and the deep research API guide covers polling, webhooks and timeouts before you choose.
