Tavily is a search API built for LLM agents: a query goes in, ranked results and extracted content come back, shaped for a model rather than for a browser.
Two of its three plans have a public price. The third has a bug where the price should be, and that is a measurement rather than a complaint.
What it costs
| Plan | Price | API credits |
|---|---|---|
| Researcher | $0 | 1,000 / month |
| Pay as you go | $0.008 / credit = $8 per 1,000 | metered |
| Project | not published | 4,000 / month |
| Enterprise | custom | custom |
⚠️ The Project plan shows a placeholder, not a price
On 2 October 2026 their pricing page rendered this where the Project price belongs:
$01234567890123456789/month
That is a digit-reel placeholder that never resolved — the kind of animated counter that is supposed to settle on a number and did not. So the Project tier has no verifiable price today, and we are not going to infer one from the credit count.
We report it because our rule for this family is that a competitor without a publicly verifiable price does not get a page — and Tavily does get one, since the free tier and the pay-as-you-go rate are both real and readable. The missing tier is named as missing instead of estimated.
$8 per 1,000 is expensive per unit, and that may be fine
Pay-as-you-go at $0.008 a credit is the fourth-highest unit rate in this comparison — behind Supadata's $16.67 entry tier and Exa's Monitors ($15) and Deep Search ($12–15) endpoints. Against ScrapingBee's $0.075 per 1,000 it is 107× more.
That gap is unit definition again, not margin. A Tavily credit buys a searched, ranked, extracted answer. A ScrapingBee credit buys one fetch of one URL that you then have to read. If you need the former, buying the latter costs you the difference in engineering time, repeatedly.
We make this point on our own behalf too: our credit buys a parsed field across eleven platforms, and comparing it to a raw-fetch credit flatters us for the wrong reason.
Who it's for
Tavily fits an agent that needs to search the web and read what it finds — research, question answering, retrieval where the corpus is "the internet".
It does not answer what gained traction in this niche over the last week, because that question needs readings taken hourly and kept. No search API answers it retroactively; the past was not photographed.
Where it is better than us
The web as a corpus. We serve eleven platforms. They search everything.
Ranking and extraction. We return what a platform published. They decide what is relevant and pull the passage out, which is a harder problem than ours.
Agent ergonomics. Their output is shaped for a model's context window. Ours is shaped for a program.
The legal part
A search API sits on crawled public content, and the compliance surface is the crawler's. We are not going to characterise their position from outside.
Ours is published: public versus personal data and a removal route.
Who measured this
We did, and we barely compete — for web search we are not an option, and a comparison that implied otherwise would be the sort this family exists to avoid.
The honest flag is the Project tier: its price was unreadable on the day we looked, and a different day may render fine. If you are deciding on that tier, the page to trust is theirs, not ours.
What we run is an hourly archive of what's rising across public platforms, which is the one question on this page a search API structurally cannot answer.
The free tier (1,000 credits/month) and the pay-as-you-go rate ($0.008/credit)
read from tavily.com/pricing on 2 October 2026 — primary source, read directly
from the page. The Project tier rendered $01234567890123456789/month, a
malformed placeholder, and is reported as having no verifiable price rather than
estimated. The $8-per-1,000 and 107× figures are our arithmetic. Verify before
deciding.