By the fastCRW team · Prices verified 2026-07-21 where marked; secondary-sourced rows are labeled · This category reprices constantly, confirm before budgeting.
Disclosure: We build fastCRW and sell one of the APIs below. Every competitor number is sourced, primary pages where possible, and the rows we could not verify first-hand say so.
Why this post exists
2026 has been the year web search APIs stopped being cheap and stopped being stable. Nebius announced an agreement to acquire Tavily in February. Exa's base search rate, cited at $5 per 1,000 in a 2025 Hacker News exchange, is $7 today. Brave replaced its free plan with a $5 monthly credit. Perplexity removed the free API credit bundled with Pro subscriptions. Parallel raised $100M at a $2B valuation selling research runs at up to $2.40 apiece. If your agent stack budget was written in 2025, it is wrong now. Here is the landscape as of July 21, 2026, in one table, with the footnotes that actually change the math.
The price table (per 1,000 search requests)
| Provider | Base price / 1,000 | The footnote that matters | Source quality |
|---|---|---|---|
| Parallel (Turbo) | $1 | ~200 ms; deeper tiers are $5, and Task API research runs go from $5 to $2,400 per 1,000 | Primary, 2026-07-21 |
| Brave (AI Grounding) | $4 + $5/M tokens | Grounded answers; free plan is now a $5/mo credit | Primary (Brave blog) |
| Perplexity Search API | $5 | Sonar models bill tokens on top; Pro subscribers lost the bundled $5 API credit in Feb 2026 | Secondary (reported) |
| Exa (Search) | $7 | Covers 10 results; each extra result +$1/1,000, deep tiers $12–$15; cited at $5 in a 2025 HN exchange | Primary, 2026-07-21 |
| Tavily | Credit-weighted | Basic and advanced searches consume different credits; acquisition by Nebius announced Feb 2026, no pricing change announced as of 2026-07-21 | Structural only |
| OpenAI web search tool | $10 + content tokens | Search content billed at model token rates on top | Primary, 2026-07-21 |
| Anthropic web search tool | $10 + content tokens | Search results count as input tokens on every subsequent turn | Primary, 2026-07-21 |
| fastCRW | $0.69 (Standard plan) | 1 credit per search, flat; failures unbilled; AGPL engine self-hosts at $0 | Primary (our pricing) |
Derivation note for our own row: the Standard plan is $69/month for 100,000 credits and a search is 1 credit, so 1,000 searches work out to $0.69; higher tiers work out cheaper, the free tier is a one-time 500 credits, and self-hosting the open-source engine has no license cost at all. Current numbers always live on the pricing page.
The model-native trap: $10 per 1,000 is the floor, not the price
Both OpenAI and Anthropic list their built-in web search tools at $10 per 1,000 searches, two to ten times the dedicated APIs. But the per-search fee is the smaller half of the real cost. Retrieved search content is billed as tokens at model rates, and on Anthropic's implementation the search results are counted as input tokens again on every subsequent conversation turn. A multi-turn agent session that searched early pays for those results on every following call. For a heavy agent product, the token side of model-native search routinely dominates the $10 fee itself.
The pattern that fixes this is boring and works: expose a dedicated search API to the model as a regular tool call (MCP or function calling), return compact result payloads, and keep raw page content out of the conversation context unless the model explicitly asks for a page. You keep citation quality and pay a fraction per query.
The free-tier squeeze is real
| Provider | Free tier, mid-2026 |
|---|---|
| Exa | $20 at signup + $10/month recurring (the most generous in the category) |
| Parallel | Advertised "up to 5,000 requests/month"; at least one public report of balance errors inside the allowance, verify on your account |
| Brave | Free plan removed; replaced by a $5/month credit |
| Perplexity | Bundled Pro API credit removed Feb 2026 |
| fastCRW | 500 credits one-time, no card; unlimited free via self-host |
The direction is uniform: venture-funded search APIs are tightening free access as they chase revenue, and 2026's funding rounds (Exa at a reported $2.2B valuation, Parallel at $2B) come with growth expectations that self-serve prices will eventually serve. Plan for the prices on this page to be the floor of what you will pay in 2027, not the ceiling.
Cost is one axis. Agents care about three.
Agent builders, and increasingly the agents themselves, evaluate search APIs on cost, latency, and reliability together, and that is the right rubric:
- Cost per 1,000: the table above, adjusted for your real results-per-request and token side effects.
- Latency: dedicated APIs run roughly 200 ms to 2 s for standard search. Our standing 100-query benchmark measured fastCRW at 880 ms average (73 of 100 latency wins), Firecrawl at 954 ms, Tavily at 2,000 ms; Parallel's Turbo advertises ~200 ms while its deeper research tiers run seconds to minutes (AIMultiple measured its Pro tier at 13.6 s average). Match the tier to the loop: an interactive agent cannot sit inside a 13-second search.
- Reliability: all three providers in our benchmark returned results on 100% of queries; the differentiator is what happens on failure. fastCRW does not bill failed requests. Check that clause in any contract you sign.
How to actually choose
- Estimate volume honestly. Under ~10,000 searches/month, everything on this page costs pocket money; optimize for quality and DX, and Exa's recurring free credit or fastCRW's self-host mode may cover you entirely.
- At 100,000+/month, the spread is real: $69 (fastCRW Standard), ~$700 (Exa base), ~$1,000 (model-native fee alone, before tokens). Run a two-week shadow test on your real queries before committing.
- Price the exit. After this year's acquisition and repricing news, treat vendor risk as a line item. An API you can self-host caps your downside at server cost; a closed index caps it at whatever the next pricing page says.
Sources
- Parallel: docs.parallel.ai pricing (2026-07-21)
- Exa: exa.ai/pricing (2026-07-21) · price history via HN · valuation as reported by Sacra
- Brave AI Grounding: brave.com/blog/ai-grounding
- OpenAI web search tool: OpenAI pricing docs (2026-07-21)
- Anthropic web search tool: Anthropic docs (2026-07-21)
- Perplexity Search API rate: reported by CloudZero's pricing breakdown; confirm on perplexity.ai
- Tavily acquisition: Nebius newsroom, 2026-02-10
Related: Exa pricing explained · What is Parallel.ai? · What the Nebius-Tavily deal means · The cheapest search API