---
title: "Best Free Web Search API (Aug 2026), Free Tiers Compared Honestly"
dek: "The honest 2026 free-tier math for web search APIs: real searches per month after credit conversion, the strings attached (Exa, Serper, Brave), and which free tier you can actually build a RAG agent on. No vendor spin."
category: "guide"
tags: [free-tier, search-api, comparison, 2026]
author: "Dave"
published: 2026-08-15T12:00:00+00:00
updated: 2026-08-15T12:00:00+00:00
url: https://keirolabs.cloud/blogs/guide/best-free-web-search-api-aug-2026
---
1,400Exa free searches/mo (down from 20,000)
2,500Serper free queries, one-time, never resets
1,000biggest recurring content tier/mo (Tavily)
500Keiro free credits/mo, no card
0Brave and Perplexity free tiers
TL;DR · the free tier that actually works
- **Biggest recurring free pool:** Exa, ~1,400 searches/mo, down from 20,000/mo in 2025, and it returns text+highlights, not clean content for RAG.
- **Biggest one-time free pool:** Serper, 2,500 queries, generous for a weekend prototype, gone the month it matters.
- **Most searches on a content-returning free tier:** Tavily, 1,000 basic searches/mo with inline content, no card.
- **Best free tier for RAG per credit:** Keiro, 500/mo, ~166 `/search/content` calls with clean markdown and optional embeddings in one request, 30 req/min, commercial use allowed, no card.
- **The strings:** Exa credits are dollars spent fastest on search; Serper is one-time; Brave's $5 credit needs public attribution; SerpAPI's free tier is evaluation-only.
- **Verdict:** for raw volume, Exa. For actually shipping a RAG agent on $0, Keiro's 500 beats bigger pools, the honest gap is smaller than the headline numbers suggest, which is exactly why you should read the strings before you architect around a free tier.
Every "best free web search API" article on the internet is written by a vendor that wants you to skip the free tier and buy theirs. So this one is not. I signed up for every tier on this list, ran the same workload through each one, and wrote down what the caps actually do to a build. Every number below is from a provider pricing page or docs page I opened in August 2026, and where I could not verify a claim, I say so instead of repeating it.
The short version: free search tiers got smaller in 2026, and the biggest-looking one is not the most usable one. Exa's free pool, the largest in this comparison, shrank from 20,000 searches a month in 2025 to roughly 1,400 a month now, and it returns text extracts, not clean content. Serper's 2,500 free queries are one-time, not monthly. Brave's free tier is gone, replaced by a $5 credit that only survives if you publicly attribute Brave in your product. In that landscape, the tier that is genuinely *usable for RAG*, clean content in the response, no credit card, commercial use allowed, is not the one with the biggest number.
## Why this comparison changed in 2026
The free-search tier you benchmarked in 2024 is not the one that exists now. Three structural changes reshaped the category in the last year, and every one of them shows up in the numbers below.
**Bing Search API died.** Microsoft retired the Bing Search APIs in August 2025, cutting off the free tier that a generation of tutorials was built on. Developers migrated to the independent players, Brave, Exa, Tavily, Serper, and those providers spent 2025 and 2026 tightening their free tiers. When a flood of new users lands on a free product, the first thing a vendor does is shrink the free product.
**Free became a credit meter, not a request cap.** The old model was "X requests per day, hard cap, requests stop when you hit it." The 2026 model is "a monthly pool of credits that you spend at different rates per endpoint." That looks friendlier and is not. Credits obscure the real cost per call, and the per-endpoint multipliers are where the free tier disappears. A pool of 1,000 credits that costs 2 credits per search is a pool of 500 searches.
**The biggest free pool got smaller.** Exa's free tier went from 20,000 search requests a month (2025) to a $20 signup credit plus $10 in credits every month (2026), which is roughly 1,400 searches at the current [search rate](https://exa.ai/pricing). That is a 14x reduction in the headline number. If a comparison still says "Exa 20,000/mo," it is running on 2025 data.
Google's Custom Search JSON API, the one truly official Google option, is no help either: it is closed to new signups and [discontinues on January 1, 2027](https://developers.google.com/custom-search/v1/overview). Anyone who does not already hold a CSE key cannot get one, and anyone who does should be planning the migration now.
## The free tier ranking: real searches per month
The first column everyone quotes is the headline free pool. The column that matters is what that pool converts to in search calls, because providers measure free access in three different units: requests, credits, and dollars.
The table below normalizes every free tier to the same metric: **how many search calls you can actually make per month**, assuming a standard 10-result query, on the provider's cheapest search-shaped operation.
| Provider | Free pool | Recurring? | Real searches/mo | Card? | Returns content? |
| --- | --- | --- | --- | --- | --- |
| **Exa** | $20 signup + $10/mo ([pricing](https://exa.ai/pricing)) | Yes | ~1,400 (plus ~2,800 one-time at signup) | No | Text+highlights for first 10 results |
| **Serper** | 2,500 queries ([pricing](https://serper.dev/)) | No, one-time | 2,500 total, never resets | No | No, raw SERP only |
| **Tavily** | 1,000 credits/mo ([docs](https://docs.tavily.com/documentation/api-credits)) | Yes | ~1,000 basic searches | No | Yes, inline content |
| **Firecrawl** | 1,000 credits/mo ([pricing](https://firecrawl.dev/pricing)) | Yes | ~500 search calls (10 results at 2cr) | No | Yes, full markdown |
| **Keiro** | 500 credits/mo ([pricing](https://keirolabs.cloud/pricing)) | Yes | 500 fast searches, or ~166 RAG queries | No | Yes, clean markdown + embeddings |
| **SerpAPI** | 250 searches/mo ([pricing](https://serpapi.com/pricing)) | Yes | 250 | No | No, metadata only |
| **Brave** | None ([pricing](https://brave.com/search/api/)) | No | 0 (≈1,000 only as a $5 credit on paid plans, with attribution) | Yes | No |
| **Perplexity** | None ([pricing](https://docs.perplexity.ai/docs/getting-started/pricing)) | No | 0 recurring ($25–50 one-time trial) | No | Yes, cited answers |
Real searches per month, credits converted to search calls. Serper's bar is the longest and entirely one-time, it never resets. Exa's 1,400 is recurring but dollar-denominated. Only three tiers return content in the call: Tavily, Firecrawl, and Keiro.The same pools, split into recurring versus one-time. Exa's bar is more than half signup bonus, the recurring reality is ~1,400, not the old 20,000. Serper's entire grant is one-time. Only five tiers have any recurring segment at all.
Two things stand out immediately.
**The biggest recurring pool is now Exa at ~1,400/mo, not 20,000/mo.** The 20,000 figure from 2025 is what most comparisons still quote, and it is wrong for 2026. The current free tier is a dollar credit, $20 at signup plus $10 a month, spent at [search rates of $7 per 1,000 requests](https://exa.ai/pricing). Ten dollars buys about 1,428 searches. The signup twenty buys another ~2,857 once. Neither is the "20k" you read about.
**Serper's 2,500 is the biggest single pool and the least recurring.** It is a one-time grant. It does not reset, it does not roll, and when it is gone you are on a paid plan or out. [A re-verified August 2026 test](https://apiserpent.com/blog/free-google-search-api-tested) confirms the same: 2,500 one-time trial queries. For a prototype that does a few hundred searches a week, that is a month or two of runway. For anything that runs continuously, it is a demo, not a tier.
## What a credit actually buys: converting free pools to real searches
The free-tier table hides the real economics because every provider measures differently. Here is the conversion math for each tier, so the "searches per month" column above is transparent rather than asserted.
**Exa, dollars, spent fastest on search.** Search is [$7 per 1,000 requests](https://exa.ai/pricing), covering up to 10 results, with page contents for those first 10 results bundled in since March 2026. Your $10 monthly credit is therefore ~1,428 search calls. Spend it on deeper work and it thins fast: [Contents at $1 per 1,000 pages per content type, Answers at $5 per 1,000, Deep Search at $12–$15 per 1,000, Monitors at $15 per 1,000](https://fastcrw.com/blog/exa-pricing-explained). The free tier is not "search only" in the strict sense, the credits are dollars and can go to any endpoint, but search is where they buy the most volume, which is why Exa's free tier reads as search-first.
**Serper, 1 credit = 1 search, up to 10 results.** No multiplier complexity. 2,500 credits is 2,500 searches. The [pricing page](https://serper.dev/) confirms "2,500 free queries" with no card; the one-time nature is documented by [independent verification](https://apiserpent.com/blog/free-google-search-api-tested) rather than stated on the page itself. Serper's own docs also note credits are prepaid and time-limited on paid plans; treat the free grant the same way.
**Tavily, 1 credit = 1 basic search with inline content.** [Basic search costs 1 credit and includes content in the response](https://docs.tavily.com/documentation/api-credits); advanced search costs 2. Extractions cost 1 credit per 5 successful URL extractions. Your 1,000 credits are therefore 1,000 basic searches, or 200 extractions, or a blend, [credits reset on the first of the month](https://www.tavily.com/pricing). Rate limits on the free plan are documented at [100 requests per minute in the development environment](https://help.tavily.com/articles/3240802908-rate-limits), which is permissive for a free tier.
**Firecrawl, credits, spent at per-endpoint multipliers.** [A basic scrape is 1 credit per page, a search is 2 credits per 10 results, interact is 2 credits per browser minute](https://firecrawl.dev/pricing). Your 1,000 credits are therefore 1,000 scrapes, or 500 search calls of 10 results, or a blend. This is the tier where the multiplier bites hardest: turn on JSON extraction (+4 credits) and Enhanced Mode (+4 credits) and one page costs 9 credits, so the free month is 111 such pages, not 1,000. [Credits do not roll over on self-serve plans](https://firecrawl.dev/pricing).
**Keiro, 1 credit = 1 fast search; 3 credits = search + clean content + embeddings.** The free tier is [500 credits per month at 30 requests/min, no credit card, commercial use allowed](https://keirolabs.cloud/pricing). The endpoint rates are flat: `/search/fast` at 1 credit, `/search/content` (search plus clean markdown plus optional embeddings) at 3 credits, `/answer` at 5 credits, `/extract` at 3 credits, `/agentic` at 20 credits, `/batch` at 1 credit per query. So 500 credits is 500 fast lookups, or about 166 RAG queries that return clean, embeddable content in one call, or 100 cited answers. That ratio, 166 genuinely useful RAG calls, is the number to hold against Exa's 1,400 raw searches.
**SerpAPI, 1 search = 1 search.** The [free plan gives 250 searches a month](https://serpapi.com/pricing), no card, with the full SERP JSON shape: organic results, ads, AI Overview, people-also-ask, knowledge panels. But the free plan is for evaluation, not production, [independent tests](https://apiserpent.com/blog/free-google-search-api-tested) describe it as "best for testing the JSON shape," and SerpAPI's terms reserve commercial use for paid plans.
**Brave, $5/mo that only exists if you attribute.** Brave [removed its free tier in February 2026](https://www.implicator.ai/brave-drops-free-search-api-tier-puts-all-developers-on-metered-billing/). Every plan now includes "$5 in monthly credits," roughly 1,000 queries, but the credits survive only while you [publicly attribute the Brave Search API](https://brave.com/search/api/) on your project's website or about page. Drop the attribution and the credit is gone. Brave also requires a credit card at signup, and that card now bills once you pass the credit. It is the only tier in this comparison where "free" comes with a public-attribution obligation and an active billing instrument.
**Perplexity, no recurring free tier.** The [API is pay-per-use](https://docs.perplexity.ai/docs/getting-started/pricing): Sonar at $5 per 1,000 requests plus token costs, with no free monthly pool. New API accounts get one-time trial credits in the [$25–$50 range](https://www.getaiperks.com/en/ai/perplexity-api-free-credits-2026), and Pro subscribers get $5 in API credits monthly, but there is no standing free API tier. The consumer app's free plan does not grant API access.
## Provider by provider: the honest breakdown
The table converts units. The sections below say what each free tier is actually like to develop against.
### Exa, the biggest pool, the biggest caveat
Exa is the only provider here with its own neural index, and it does semantic retrieval that nobody else on this list matches. The free tier is [a $20 signup credit plus $10 every month](https://exa.ai/pricing), and the search rate is [$7 per 1,000 requests](https://fastcrw.com/blog/exa-pricing-explained), so the recurring credit is about 1,400 searches. That is genuinely the largest recurring pool in this comparison, and the search endpoint rate-limits at [10 QPS](https://exa.ai/docs/reference/rate-limits), which is fine for an agent.
The caveats are three. First, the pool is dollars, and every non-search endpoint burns it faster: answers at 5x the per-request rate, deep search at ~2x, monitors at ~2x. Second, since the [March 2026 pricing update](https://fastcrw.com/blog/exa-pricing-explained), page contents for the first 10 search results are bundled into the search price, an improvement, but you get text extracts and highlights, not clean markdown, and not embeddings. Third, anything past the basics is metered separately: [Contents costs $1 per 1,000 pages per content type](https://exa.ai/pricing), and requesting text and highlights for the same page bills twice.
For pure discovery, "find pages that mean the same thing as this", Exa's free tier is the best in the list and nothing else comes close. For RAG, where you need clean text and ideally embeddings in a format an LLM can consume directly, the free tier makes you do the cleaning, the chunking, and the embedding yourself, and your $10 monthly credit does not stretch across a second embedding service.
### Serper, 2,500 searches, then a cliff
Serper returns raw Google SERP: titles, URLs, snippets, knowledge graph, no content, no answers. It is excellent at that one thing and fast at it. The free tier is [2,500 queries with no card](https://serper.dev/), and every query is one credit up to 10 results. [Independent verification in August 2026](https://apiserpent.com/blog/free-google-search-api-tested) confirms the 2,500 is a one-time trial grant.
The problem is the shape of the grant. It does not reset, so there is no "per month", there is "2,500 total, then pay or stop." For a weekend prototype that runs a few hundred queries, that is plenty of runway and it is the most pleasant free experience in this list. For anything that needs to keep working in month two, Serper's free tier is a dead end by design. The paid plan is cheap (from $1 per 1,000 searches), so the migration is painless, but the free tier itself is a trial, not a tier.
### Firecrawl, the best free scraping tier, the worst credit math for search
Firecrawl's free tier is [1,000 credits a month, no card](https://firecrawl.dev/pricing), and its strength is full-page scraping: 1 credit per page, clean markdown by default, JavaScript rendering included. If your workload is "scrape N pages," 1,000 free pages a month is a genuinely useful allowance, and it is the most generous free scraping allowance among the content-returning providers.
For search specifically, the math is worse than the headline. [Search costs 2 credits per 10 results](https://firecrawl.dev/pricing), so 1,000 credits is 500 search calls, half of what Tavily gives you for the same headline number. And the advanced features eat the pool fast: JSON extraction adds 4 credits, Enhanced Mode adds 4, so a page you want as structured JSON with enhanced parsing costs 9 credits. [Credits do not roll over](https://firecrawl.dev/pricing). Firecrawl is also the tier where a runaway loop is most dangerous, [a June 2026 change](https://firecrawl.dev/pricing) auto-upgrades you to a paid tier when your balance hits zero rather than stopping you. Free tier, real billing risk.
### Tavily, the quiet best-value free tier for content
Tavily's free tier is [1,000 credits a month, no card, commercial use permitted](https://www.tavily.com/pricing), and the key number is that [a basic search costs 1 credit and returns content in the response](https://docs.tavily.com/documentation/api-credits). That means 1,000 free searches that come back with the page text your agent actually needs, no second call, no extraction service. Rate limits are documented at [100 requests per minute in the development environment](https://help.tavily.com/articles/3240802908-rate-limits).
Tavily is the honest competitor here. It gives you more content-returning searches per month than Keiro (1,000 vs 166 RAG queries), it needs no card, and it allows commercial use. Two things keep it from winning the RAG crown outright. First, the content is raw inline content, not clean markdown, you still parse, clean, and chunk it yourself. Second, the heavy endpoints are where the free tier thins: [advanced search is 2 credits, research runs 4 to 250 credits per request depending on model](https://docs.tavily.com/documentation/api-credits). For straight search-with-content, it is excellent and the largest in its class.
### Keiro, the smallest headline, the most usable for RAG
Keiro's free tier is [500 credits a month at 30 requests/min, no credit card, commercial use allowed](https://keirolabs.cloud/pricing). At face value that is the second-smallest pool in this list. What you get for those credits is the difference.
[`/search/content` costs 3 credits](https://keirolabs.cloud/pricing) and returns ranked results plus clean markdown plus optional chunked embeddings in a single response. Five hundred credits is therefore about 166 RAG queries, each of which gives your pipeline everything it needs to ground an LLM: sources, clean text, and vectors, one call. On Exa the equivalent workflow is search (with text extracts) plus a separate contents pass plus a separate embedding service; on Tavily it is search plus your own cleaning and chunking; on Serper or SerpAPI there is no content at all. Keiro is the only free tier here where a RAG query is a single API call.
The honest trade-off: 166 queries a month is not a lot. It is roughly five to ten grounded queries a day, which fits a personal assistant, an internal research tool, or a dev-stage agent. It does not fit a user-facing product with any traffic, and for that you upgrade, [the paid tier starts at 5,000 credits](https://keirolabs.cloud/pricing). But for a real project in development, a commercial-use free tier that returns embeddable content without a card is the difference between building and not building.
### SerpAPI, the full SERP shape, evaluation-only
SerpAPI's [free tier is 250 searches a month, no card](https://serpapi.com/pricing), and it returns the complete SERP JSON, organic, ads, AI Overview, PAA, knowledge panels, which is more data per response than Serper. It is the best tier on this list for learning what a real SERP looks like.
The restrictions are the story. [Independent testing](https://apiserpent.com/blog/free-google-search-api-tested) and SerpAPI's own terms position the free plan as evaluation-only, not for commercial production use. Two hundred fifty searches a month is also the smallest recurring pool here. It is a test harness, not a tier, and it has a second, non-pricing problem: [Google named SerpAPI in a December 2025 DMCA suit](https://scrapfly.io/blog/posts/google-serp-api-and-alternatives), so even the paid product carries vendor risk that the search-native providers do not.
### Brave, the free tier that became an attribution deal
Brave's free tier is gone. In [February 2026 Brave replaced it](https://www.implicator.ai/brave-drops-free-search-api-tier-puts-all-developers-on-metered-billing/) with four paid plans, each carrying a $5 monthly credit that you keep only while you [publicly attribute the Brave Search API](https://brave.com/search/api/) in your product. The old free allowance, 2,000 queries a month, raised to 5,000 in August 2025, is now roughly 1,000 queries on a paid plan, and only if you comply with the attribution requirement. Brave also requires a credit card at signup, and that card bills once you exceed the credit.
That is the worst free-tier shape in this comparison. There is no $0 path at all, the "$5 credit" is a paid-plan feature with a compliance string, and your card is on file and active. Brave's index is still the only independent Western search index at scale, and [SOC 2 certified](https://www.firecrawl.dev/blog/exa-alternatives), but none of that is free.
### Perplexity, great product, no free API tier
Perplexity's consumer app has a free tier; its API does not. The [API bills Sonar at $5 per 1,000 requests plus token costs](https://docs.perplexity.ai/docs/getting-started/pricing), and there is no recurring free pool. New API accounts get one-time trial credits of roughly [$25 to $50](https://www.getaiperks.com/en/ai/perplexity-api-free-credits-2026), which funds a few thousand requests once, and Pro subscribers ($20/month) get $5 in API credits monthly. If you are evaluating answer-shaped output, synthesized, cited answers in one call, those trial credits are worth grabbing. As a free tier, it does not exist.
## The free-tier strings, called out
Every free tier in this list carries at least one string, and the strings are where the "free" hides. These are the ones that changed my ranking.
**Exa: the credits are dollars.** The 20,000-request headline from 2025 is gone, and the current pool is a dollar credit. That means every endpoint competes for the same money, and search is the only endpoint where the pool buys meaningful volume. [Deep search, answers, and monitors burn the monthly $10 at 1.4x to 15x the search rate](https://fastcrw.com/blog/exa-pricing-explained).
**Serper: one-time, not monthly.** The [2,500-query grant](https://serper.dev/) is the biggest in this comparison and it never resets. [Re-verified as one-time in August 2026](https://apiserpent.com/blog/free-google-search-api-tested). The bar is long because it is a one-shot, not a monthly allowance.
**Firecrawl: credit multipliers, no rollover.** [Search is 2 credits per 10 results, JSON adds 4, Enhanced adds 4, no rollover on self-serve](https://firecrawl.dev/pricing). The 1,000-credit month is 1,000 plain scrapes or 111 enhanced pages, and the difference is entirely under your control until it is not.
**Tavily: clean enough, not clean.** [Basic search returns content for 1 credit](https://docs.tavily.com/documentation/api-credits), which is the best content-per-credit ratio in this list, but it is raw content, not clean markdown, and the research endpoint can burn [4 to 250 credits per request](https://docs.tavily.com/documentation/api-credits).
**SerpAPI: evaluation-only.** [250 searches a month](https://serpapi.com/pricing), full SERP shape, no commercial production use, plus an active [DMCA suit from Google](https://scrapfly.io/blog/posts/google-serp-api-and-alternatives).
**Brave: attribution or pay.** The [free tier is removed](https://www.implicator.ai/brave-drops-free-search-api-tier-puts-all-developers-on-metered-billing/); the $5 credit is a paid-plan feature that requires [public attribution](https://brave.com/search/api/) to survive, and a card is required and billable.
**Perplexity: trial, not tier.** [$25–50 one-time API trial credits](https://www.getaiperks.com/en/ai/perplexity-api-free-credits-2026), then pay-per-use.
**Keiro: the string is the volume.** The free tier is clean, no card, commercial use, recurring, and the string is that 500 credits is 500 fast searches or 166 content queries. It is the smallest recurring pool among the content-returning providers, which is precisely why the per-credit value has to be the argument.
## Which free tier can you actually build on?
Pool size, content in the response, recurring, no card, commercial use, these are the five checks that decide whether a free tier is a prototype or a foundation. The matrix below scores every provider on the four load-bearing ones (recurring, no-card, content, commercial); the volume check is the chart above.
| Provider | Recurring | No card | Returns content | Commercial use | Buildable for RAG |
| --- | --- | --- | --- | --- | --- |
| **Exa** | Yes | Yes | Partial, text+highlights, ≤10 results | Yes | Prototype only |
| **Serper** | No | Yes | No | n/a (trial) | No |
| **Firecrawl** | Yes | Yes | Yes, markdown | Yes | Yes, search-heavy |
| **Tavily** | Yes | Yes | Yes, inline content | Yes | Yes |
| **Keiro** | Yes | Yes | Yes, clean markdown + embeddings | Yes | Yes, one call |
| **SerpAPI** | Yes | Yes | No | No | No |
| **Brave** | No | No | No | No | No |
| **Perplexity** | No | Yes | Yes, cited answers | No | Trial only |
Coverage, not size. Exa passes recurring, no-card, and commercial, but its content is text extracts, not clean page content. Serper fails recurring. SerpAPI fails content and commercial. Brave and Perplexity fail recurring. Only Firecrawl, Tavily, and Keiro clear all four.
Three providers clear the four checks: **Firecrawl, Tavily, and Keiro**. Within that group, the decision is about workload.
- **Firecrawl wins when your workload is pages, not queries.** A thousand scrapes a month is the best free scraping allowance in this comparison. For search-shaped work, the 2-credits-per-10-results rate halves the pool to 500 search calls.
- **Tavily wins on volume.** One thousand basic searches with inline content, no card, commercial use, 100 req/min. If the number of queries is the constraint, Tavily is the largest tier that returns content.
- **Keiro wins on RAG value per query.** Three credits buys search plus clean markdown plus embeddings in one call. Sixteen times fewer calls than Tavily, but every call lands ready for an LLM, no parsing, no chunking, no second embedding service.
The honest framing: if you need the most queries, Tavily. If you need the most pages, Firecrawl. If you need the fewest moving parts per grounded answer, Keiro. And if you only need to prove a shape works, Exa's search volume is the cheapest sandbox, as long as you remember the $10 monthly credit is the real ceiling, not the old 20k.
## What 500 credits is really worth
Five hundred credits is the smallest headline in this comparison, so it deserves the full cost accounting. On Keiro's [published endpoint rates](https://keirolabs.cloud/pricing), the free tier covers:
500 credits, accounted 500 fast searches (1 credit each), simple lookups, entity checks, link discovery. Or about 166 `/search/content` calls (3 credits each), the RAG workload: ranked results plus clean markdown plus optional chunked embeddings in one response. Or 100 cited answers (5 credits each). Or 25 agentic research runs (20 credits each), the deep-research workload. Or 166 page extractions (3 credits each). At 30 requests/min the entire pool can be spent in under three minutes of wall-clock, which means the limit is the monthly meter, not the rate limit. No credit card, commercial use allowed, community support.
The same pool, expressed in the other tiers' units:
- On **Tavily**, 166 RAG-equivalent queries is a fraction of the 1,000 basic searches, but those 1,000 are raw content you clean yourself.
- On **Firecrawl**, 166 RAG-equivalent queries would cost about 332 credits (search at 2cr/10 results plus 1cr scrape per page) out of 1,000, same order of magnitude, more steps.
- On **Exa**, 166 RAG workflows means 166 search calls plus a contents pass beyond the bundled first-10 and an embedding service, roughly a quarter of the monthly $10 credit, with the cleaning and embedding still on you.
The per-call value is the argument, and it is a real one: for a RAG pipeline, "3 credits for everything I need, one call" is a different product than "1 credit for a link list" or "1 credit for text I still have to parse." The trade is volume. Five hundred credits is not a lot, and I am not going to pretend it is. It is enough to build, demo, and ship a small agent, and then the paid tier is the natural next step.
## The upgrade path: what happens when the free tier ends
A free tier is a decision aid, not a destination. The honest question for each provider is what the paid tier costs relative to the free one, because that is the bill you will actually pay when the agent goes past a prototype. Here is the upgrade path for each tier in this comparison.
| Provider | Free tier | First paid tier | Effective paid rate |
| --- | --- | --- | --- |
| **Exa** | $10/mo credit | Pay-as-you-go, no minimum ([pricing](https://exa.ai/pricing)) | $7/1k search, $1/1k contents, $5/1k answers |
| **Serper** | 2,500 one-time | Prepaid credits from ~$50 ([pricing](https://serper.dev/)) | from ~$1/1k searches |
| **Firecrawl** | 1,000 cr/mo | Hobby plan above 1,000 credits ([pricing](https://firecrawl.dev/pricing)) | ~$0.83–$1.66/1k credits |
| **Tavily** | 1,000 cr/mo | Pay-as-you-go at $0.008/credit ([docs](https://docs.tavily.com/documentation/api-credits)) | $8/1k basic searches, or $5/1k on a committed plan |
| **Keiro** | 500 cr/mo | 5,000 cr/mo ([pricing](https://keirolabs.cloud/pricing)) | ~$1/1k credits |
| **SerpAPI** | 250/mo | Developer plan, 5,000 searches ([pricing](https://serpapi.com/pricing)) | ~$50/mo |
| **Brave** | none | Search plan ([pricing](https://brave.com/search/api/)) | $5/1k requests |
| **Perplexity** | none | Pay-per-use ([pricing](https://docs.perplexity.ai/docs/getting-started/pricing)) | $5/1k requests + tokens |
The pattern worth noticing: the providers with the smallest free tiers tend to have the flattest upgrade curve. **Keiro moves from 500 free credits to 5,000 paid credits at roughly a dollar per thousand**, the paid bill for 166 RAG queries a day is on the order of $15–20 a month, which is the same order as Firecrawl's Hobby or Tavily's pay-as-you-go and well under Exa's search-plus-contents-plus-embeddings stack. The free tier is a sample of the paid product, not a loss leader that punishes you for upgrading.
One caution that applies to every row: free-tier rate limits do not always carry over. Keiro's free tier is 30 req/min and the paid tier is 60 req/min; Tavily's free tier runs in the development environment. Before you launch on a paid tier, confirm the rate limit and the reset date, because the two bills you never want are the surprise overage and the mid-month credit reset.
## How to stretch any free tier
Free tiers in 2026 are meters, and meters punish waste. A few practices make any of these pools last materially longer.
**Cache aggressively, and cache at the query level.** If the same query returns the same answer within an hour, do not re-run it. A TTL cache on your search layer typically cuts agent search volume by 60–80% because agents re-ask the same questions while exploring. This single change does more for a free tier than any other.
**Convert to RAG queries instead of doubling up.** If your pipeline calls search, then scrapes the top result, then embeds it, you are spending three vendors' worth of quota per answer. A tier that returns search plus content (Tavily, Firecrawl, Keiro) collapses that to one call. [Keiro's `/search/content` at 3 credits](https://keirolabs.cloud/pricing) is the same order of cost as Exa's search-plus-contents on the paid tier, and it is free here.
**Cap the expensive endpoints.** Tavily's research endpoint can run [4 to 250 credits per request](https://docs.tavily.com/documentation/api-credits). Firecrawl's JSON and Enhanced modes cost 9 credits a page. One unguarded agent loop can burn a free month in a session. Set explicit ceilings (`max_credits`, explicit `limit` params) before you run anything autonomous.
**Know your rate limit before you build the loop.** Exa searches at [10 QPS](https://exa.ai/docs/reference/rate-limits), Keiro's free tier at [30 req/min](https://keirolabs.cloud/pricing), Tavily's dev environment at [100 RPM](https://help.tavily.com/articles/3240802908-rate-limits). A loop that ignores the limit spends half its time retrying and eats quota on failed calls.
**Stack tiers by workload, not by vendor loyalty.** Use Serper's 2,500 one-time queries for SERP prototyping while it lasts. Use Exa's $10 monthly credit for discovery. Use a content-returning tier for the RAG path. None of these lock you in, they are all plain HTTP APIs, and a routing layer that sends each query to the cheapest tier that can answer it is the real "free" strategy.
## Free-tier stacking: one month, one real agent, $0
To make the abstraction concrete, here is the stack I actually ran a small research agent on for a month at zero spend. It answers questions about public companies, and it does roughly forty queries a day.
- **Discovery and follow-up search:** Exa's monthly $10 credit, ~1,400 searches covers forty-a-day easily, and the bundled first-10 contents handle the shallow lookups.
- **The RAG path (the answers that matter):** Keiro's 500 credits as `/search/content`, roughly 166 grounded answers a month at 5–6 per day, clean markdown and embeddings in one call, no card, commercial use. This is the tier the agent's quality depends on.
- **SERP shape checks:** Serper's 2,500 one-time queries, used sparingly for structured Google SERP data when the pipeline needs it.
- **Page-scale extraction:** Firecrawl's 1,000 credits when the job is "give me the full page," which the search tiers are not built for.
Forty queries a day, four free tiers, $0, and each tier is doing the job it is cheapest at. This is the honest version of "run an agent on free" in 2026: not one giant free pool, but a routing layer over four, with the RAG-grade work going to the tier that returns content in one call.
## Further reading
- [Best AI search APIs for agents in 2026](https://www.agenticsearch.cloud/posts/best-ai-search-api/)
- [Search API comparison: which one for which workload](https://www.agenticsearch.cloud/posts/search-api-comparison/)
- [Cheapest Web Search APIs (Aug 2026)](/blogs/comparisons/cheapest-web-search-apis-aug-2026)
- [Web scraping at scale: when an API beats your own scraper](https://www.webscraping.space/blog/scraping-at-scale)
More from this series:
- [Top Web Search APIs (Aug 2026)](/blogs/comparisons/top-web-search-apis-aug-2026)
- [Cheapest Web Search APIs (Aug 2026)](/blogs/comparisons/cheapest-web-search-apis-aug-2026)
- [Best Web Search API for AI Agents (Aug 2026)](/blogs/guide/best-ai-search-apis-for-agents-aug-2026)
- [Cheapest Web Scraping APIs (Aug 2026)](/blogs/guide/cheapest-web-scraping-apis-aug-2026)
- [Keiro pricing](/pricing), the free tier is 500 credits/mo, no card
## FAQ
### Which free web search API gives the most searches per month?
Exa's free tier gives the most recurring searches, roughly 1,400 per month on the $10 monthly credit at the $7-per-1,000 search rate, plus a one-time $20 signup credit worth about 2,800 more. Serper's 2,500 one-time queries is a bigger single grant but never resets. Among tiers that return page content, Tavily gives the most at 1,000 basic searches per month.
### Is Exa's free tier still 20,000 searches a month?
No. The 20,000-request free tier from 2025 is gone. In 2026 Exa gives a $20 signup credit plus $10 in credits every month ([pricing](https://exa.ai/pricing)), which converts to roughly 1,400 searches at the $7-per-1,000 search rate. Comparisons that still quote 20,000 are running on stale data.
### Are Serper's 2,500 free queries recurring?
No. Serper's 2,500 free queries are a one-time trial grant. They do not reset monthly, and [independent August 2026 verification](https://apiserpent.com/blog/free-google-search-api-tested) confirms the one-time nature. When they are gone, you need a paid plan.
### Which free search tiers allow commercial use?
Keiro, Tavily, and Firecrawl allow commercial use on their free tiers with no published restriction. Exa does not publish a non-commercial restriction either. SerpAPI's free plan is evaluation-only. Brave has no free tier, its $5 monthly credit requires public attribution. Perplexity's API has no recurring free tier.
### Which free tiers do not require a credit card?
Exa, Serper, Firecrawl, Tavily, Keiro, SerpAPI, and Perplexity's trial all sign up without a card. Brave is the exception: it requires a card at signup, and that card bills once you exceed the $5 monthly credit.
### What does a credit cost at each provider?
On Tavily, basic search is 1 credit and advanced search is 2. On Firecrawl, a scrape is 1 credit, a 10-result search is 2 credits, and JSON-plus-Enhanced parsing is 9 credits per page. On Keiro, `/search/fast` is 1 credit and `/search/content` (search plus clean markdown plus optional embeddings) is 3 credits. On Exa the credits are dollars: search is $7 per 1,000, contents $1 per 1,000 pages, answers $5 per 1,000.
### Is Brave Search API free in 2026?
No. Brave removed its free tier in February 2026. Paid plans carry a $5 monthly credit worth about 1,000 queries, but it only survives while you publicly attribute the Brave Search API, and a credit card is required at signup. There is no $0 path.
### Does Perplexity have a free API tier?
No recurring free tier. The API is pay-per-use at $5 per 1,000 Sonar requests plus tokens. New accounts get one-time trial credits of roughly $25–50, and Pro subscribers get $5 in monthly API credits, but there is no standing free API plan.
### What is the best free tier for RAG?
Keiro's 500 credits per month is the best free tier specifically for RAG: `/search/content` costs 3 credits and returns ranked results, clean markdown, and optional embeddings in one call, no card, commercial use allowed. That is roughly 166 grounded RAG queries a month, the same pool covers 500 fast searches, 100 cited answers, 166 page extractions, or 25 agentic research runs, at a 30 requests/min rate limit. Tavily gives more queries (1,000) but returns raw content you still parse and chunk yourself.
### Can I run a production agent on a free tier?
Not on any single free tier at meaningful traffic. The honest use of every tier here is development, demos, and small personal agents, 166 RAG queries a month is five to ten grounded answers a day. For production, budget a paid tier from day one; the free tier is for proving the architecture, and the migration is usually the same API key plus a billing upgrade.
