By the fastCRW team · Every price below read directly off exa.ai on 2026-07-21 · Vendor pricing is volatile, confirm on exa.ai/pricing before deciding.
Disclosure: We build fastCRW, a search and scraping API, so we are not neutral about Exa. Every Exa number here comes from their live pricing page, docs, or changelog as seen on 2026-07-21, and where their model is a good deal we say so.
The short version
Exa sells API access to its own neural web index. Pricing is pay-as-you-go with no monthly minimum, which is genuinely developer-friendly. The complexity, and the surprise bills, come from three places: per-content-type billing on the Contents endpoint, a $1 per 1,000 surcharge for every result past the tenth on nearly every endpoint, and the new Agent API whose default mode has a variable cost spanning nearly two orders of magnitude. Base search also rose from $5 to $7 per 1,000 requests between 2025 and 2026. Here is the whole model, line by line.
The endpoint price table (verified 2026-07-21)
Base price covers up to 10 results per request; every additional result bills $1 per 1,000 on top:
| Endpoint | Base / 1,000 requests | Extra results >10 |
|---|---|---|
| Search | $7 | $1 / 1,000 |
| Deep Search | $12 | $1 / 1,000 |
| Deep-Reasoning Search | $15 | $1 / 1,000 |
| Answer | $5 | $1 / 1,000 |
| Monitors | $15 | $1 / 1,000 |
| Contents | $1 per 1,000 pages per content type | n/a |
AI summaries add another $1 per 1,000 pages wherever you enable them. Two structural notes. First, the old neural/keyword split is gone: the current search types are instant, fast, auto, deep-lite, deep, and deep-reasoning, and everything below the deep tiers bills at the single $7 rate. Second, since a March 2026 pricing update, page contents for the first 10 search results are bundled into the $7, a real improvement over the old model where contents always billed separately.
The free tier is better than most
Credit where due: Exa grants $20 of credits at signup plus $10 of credits every month, with full API access. Unlike the one-time-grant pattern common in this category (Firecrawl's 1,000 one-time credits, for instance), the monthly $10 keeps hobby projects alive indefinitely, roughly 1,400 basic searches a month for free. There is also an unauthenticated MCP free tier (3 QPS, 150 calls/day) and a $1,000 startup/education grant program. If you are evaluating Exa, the free tier will not be the thing that stops you.
Trap 1: Contents bills per content type, per page
The Contents endpoint charges $1 per 1,000 pages per content type. Request text and highlights for the same page and you are billed twice; add summaries and it is three times. Teams read "$1 per 1,000 pages" and budget one line item, then discover their actual multiplier in the invoice. If you only need clean page text for RAG, request exactly one content type, or better, stay inside the first 10 search results where contents are bundled.
Trap 2: the eleventh result costs extra, everywhere
Every endpoint's base price covers 10 results. Ask for 30 results per search, a completely normal retrieval setting for RAG re-ranking, and each request carries 20 extra-result charges. At $1 per 1,000 extra results, a 30-result search effectively bills $7 + $20 = $27 per 1,000 requests, nearly four times the headline rate. Larger per-request result counts run into Enterprise-only territory: Exa's docs gate the highest limits (up to 1,000 results per search) behind "contact sales". If your retrieval depth is large, price that, not the $7.
Trap 3: Agent's default effort mode is a cost lottery
The Agent API (launched June 2026) prices along four dimensions at once: compute units at $0.10 per ACU, search calls at $0.005 each, email enrichment at $0.02, phone enrichment at $0.07. The fixed effort modes span $0.012 (minimal) to $1.00 (x-high) per request, and the default effort: auto floats inside that range based on what the agent decides to do. Contact enrichment bills on top even in fixed modes. This is the same pattern we flag in every deep-research product: per-task pricing where the task decides its own depth is unforecastable by construction. If you use Agent in production, pin an effort mode and instrument per-request cost from day one.
Websets is a separate subscription
Websets, Exa's list-building product, does not draw from API pay-as-you-go. It has its own plans: Free (1,000 credits, capped features), Core at $49/month for 8,000 credits, Pro at $449/month for 100,000 credits, Enterprise custom. A verified "all-green" result consumes 10 credits, so Core buys roughly 800 verified rows a month, and enrichment columns and alerts consume credits on top. If your GTM team adopts Websets while engineering runs the API, you are running two Exa bills with different meters.
What Exa actually costs at realistic volumes
- Prototype agent, 5,000 searches/mo, 10 results: $35/mo, minus the $10 monthly free credit = ~$25/mo. Cheap. This is the workload Exa's model is built for.
- Production RAG, 100,000 searches/mo at 30 results: 100k × ($7 + $20)/1,000 = ~$2,700/mo. The extra-results surcharge, not the base rate, is the bill.
- Deep-reasoning research loop, 20,000 requests/mo: 20k × $15/1,000 = $300/mo before contents and summaries.
And the price direction has been up: in a public Hacker News exchange, a commenter identifying as Exa staff cited "$5 per 1,000 requests" for base search; the same call is $7 today, a 40% increase, and developers in that thread were already saying the economics did not work for grounding at volume.
The structural comparison
Exa's index is real technology: embedding-first retrieval over its own crawl, with sub-second instant search. If semantic retrieval over a curated index is the product you need, Exa is a legitimate choice and the free tier is generous. The structural costs are the same as every closed index: no self-hosting, no cost ceiling you control, and pricing that has moved upward.
fastCRW makes the opposite bet: live search plus scraping in one engine, flat metering, and an open-core escape hatch. A search is 1 credit, a scraped page is 1 credit, on every renderer, and failed requests bill nothing. On the Standard plan ($69/mo for 100,000 credits as of this writing; current numbers always live on the pricing page) 1,000 searches work out to $0.69, about a tenth of Exa's base rate, and the AGPL-3.0 engine self-hosts for free when you want the marginal cost to be zero. Different retrieval model, to be clear: we hit the live web rather than a proprietary embedding index, so compare quality on your own queries. One public data point: on ArXivQA research-paper retrieval (191 questions, recall-scored), fastCRW measured 61.0% recall versus Exa's 43.4%; methodology on the benchmark page.
How to forecast your real Exa bill
- Fix your results-per-request first. The bill scales with it: 10 results = $7, 30 results = $27 effective per 1,000 searches.
- Count content types, not pages. Text + highlights + summaries = 3 line items per page on Contents.
- Pin Agent effort modes. Never ship
effort: autowithout per-request cost telemetry. - Budget Websets separately. It is a second subscription with its own credit math.
- Price the alternative. Whatever your monthly Exa number is, compare it against a flat-metered API and against self-hosting an open engine. The delta is your negotiating leverage.
Sources
- Exa pricing: exa.ai/pricing (verified 2026-07-21)
- Search types and API reference: exa.ai/docs/reference/search · changelog: exa.ai/docs/changelog
- Websets plans: exa.ai/docs/websets/faq
- The $5-to-$7 change: Hacker News thread (a commenter identifying as Exa staff citing the earlier $5 rate)
Related: What is Exa AI? · Best Exa alternatives · Web search API pricing in 2026