Answer and research APIs
We asked 5 answer APIs the same 45 real agent questions and judged every answer against a key verified on the primary source. Parallel Task API (lite) is the most accurate but takes about a minute per answer; Parallel Responses is the most accurate for the time. Every one of them is sometimes wrong, and the wrong answers read as confidently as the right ones.
Answer APIs, judged
The same 45 real agent questions as the search eval, sent verbatim, one call each. Every answer was checked against a key verified on the primary source and judged correct, partly right, wrong or declined (the answer says it can't tell). Ordered by correct answers; ties are broken by fewer wrong answers.
- 1 Parallel Task APIlite processor, text output 41/45
- 2 Parallel Responseslow reasoning effort 40/45
- 3 Exa /answerone cited paragraph 37/45
- 4 LinkupsourcedAnswer, standard depth 35/45
- 5 Tavilysearch with advanced answer 35/45
- correct
- partly right
- wrong
- declined
- of 45 questions; median time per answer; list price per 1,000 calls
- Parallel Task API (lite) is the most accurate, 41/45, but takes about a minute per answer, so it fits a background lookup rather than an agent waiting on the reply. Parallel Responses (low reasoning effort) is the best accuracy for the time: 40/45 in 2.7 s.
- On one run, a gap of 2 or 3 correct answers is not a firm ranking. The 6-answer gap between the top and the bottom is likelier to hold.
- Linkup was the only API that sometimes said it couldn't tell (2 times), which is better than guessing.
Confidently wrong
Every API was wrong sometimes, and the wrong answers read exactly as confidently as the right ones. Some examples, described without giving the questions away. Open the cited page before acting on an answer.
- Tavily named an older release, with a date two weeks off, as a tool's latest stable version.
- Tavily told an app to add protected fields straight into its API calls, when they must be requested through a dashboard.
- Parallel Responses gave a flat "no" to a store-policy question that has a documented exception.
- Parallel's Task API (lite) gave a GPU per-second price derived from an hourly rate, about 10% above the published per-second price, and stated it as exact.
- On a home-networking question, Exa, Parallel Responses, Linkup and Tavily all described a "dedicated backhaul" feature that the product's documentation doesn't describe.
How the answers were judged
- First, string matching: an answer containing every required fact counts as correct (Parallel Task API 17, Parallel Responses 30, Exa /answer 22, Linkup 15, Tavily 11).
- Then every answer that missed was read by hand against the key (130 in all) and judged with one bar for every API. A failed call would score zero; none failed.
- String matching undercounts Tavily in particular: its answers write hyphens as U+2011 (a non-breaking hyphen), so "10‑ms" doesn't match "10-ms". The hand judging corrects for this, so compare the APIs on correct answers, not on string matches.
- One run, judged by one reader. Prices are list prices per 1,000 calls on the measurement date; median time is per answer.
- Not run: Jina DeepSearch, because one probe question used about 271k tokens and 82 s, so 45 questions would cost about twice the free token balance.
- Not run: You.com, because the account's free credit read as depleted.
Tools
Alphabetical. Links go to each vendor's own site.
| Tool | What it does | Evidence | Price or terms |
|---|---|---|---|
| Brave Answers | cited answers from Brave's own search index | none yet | $4/1k + tokens |
| Exa /answer | one cited paragraph per question | 37/45 correct, 3 wrong (#3 of 5) | $5/1k; Deep Search $12/1k |
| Linkup Research | multi-step research runs with citations | none yet | $250–2,500/1k |
| Linkup sourcedAnswer | a short cited answer from Linkup search | 35/45 correct, 2 wrong (#4 of 5) | $6/1k at standard depth |
| Parallel Responses | chat-style cited answers with a reasoning-effort setting | 40/45 correct, 2 wrong (#2 of 5) | $10/1k at low effort |
| Parallel Task | long-running research tasks, priced by processor tier | 41/45 correct, 1 wrong (#1 of 5) | $5–2,400/1k runs |
| Perplexity Sonar | cited answers from Perplexity's search, billed by tokens | none yet | sonar $0.25 / $2.50 per 1M tokens |
| Tavily Research | multi-step research reports, billed in credits | none yet | 4–250 credits per request |
| Tavily search with answer | Tavily search returning a generated answer beside the results | 35/45 correct, 4 wrong (#5 of 5) | 1 credit per call |
| You.com Research API | research answers in tiers from Lite to Frontier | none yet | Lite $12 to Frontier $1,200 per 1k |
Caveats
- Prices change often; catalog prices were compiled on the catalog date and not individually re-verified. Measured prices are list prices on the measurement date.