Parallel- [Best AI search for agents: 6 web search APIs compared (2026)](https://parallel.ai/articles/best-ai-search-for-agents)
Tags:Comparison
October 29, 2026
# The best free web search APIs for AI agents in 2026
Free tiers for web search APIs don’t line up. Some are one-time credit grants, some refill monthly, some need a card, and rate limits usually bind before credits do. This guide compares the free tiers and rate limits of 12 providers as of October 27, 2026, and matches each one to prototyping, hobby, or production use.
Tags:Comparison
## **Which free web search API should you use?**
For an AI agent, start with one of these five. Parallel’s hosted MCP server gives any MCP client web search and page fetch with no account, key, or card, and a Parallel API account with a card on file adds $5 in credits each month, enough for up to 5,000 Turbo or Fast searches. Tavily (1,000 credits a month), Exa ($10 a month plus a $10 signup bonus), and Linkup (topped back up to $20 a month) need no card at all. TinyFish makes search free up to its rate limits. Google’s free Custom Search JSON API is closed to new customers, and Microsoft retired the Bing Search APIs in August 2025.
Every free-tier figure below was checked against the vendor’s own pricing or documentation page on October 27, 2026, and each one links to its source.
## **What “free” actually means for a search API**
“Free tier” covers at least four different arrangements, and which one you’re offered matters more than the headline number.
### **Are the credits one-time or recurring?**
A signup grant is a fixed amount you use up once. A recurring monthly allowance refills, which is what you want for a personal agent or a side project that runs indefinitely. Several providers offer both, and pricing pages often blur the two. Most recurring allowances expire at month end rather than rolling over, so unused credit is gone.
### **Is a card required?**
A hard-capped free tier with no card on file can’t generate a bill. A credit-based free tier with a card attached can convert to metered billing the moment you pass the credit, and not every provider offers a spending cap that blocks requests. Parallel is one that doesn’t. Its monthly spend limits[monthly spend limits] email admins but never block requests. The two behave differently if your agent enters a retry loop at 3 a.m.
### **Which runs out first, credits or rate limits?**
Free tiers are usually throttled well below paid tiers. A plan with 1,000 monthly credits and two concurrent requests will bottleneck on concurrency long before you spend the credits. If you’re testing latency or building anything with parallel fan-out, check the rate limit before the credit figure. The table below lists both.
### **Where do the results come from?**
Some providers run their own web index. Others reformat a consumer search engine’s results as JSON. That SERP data is cheap and often excellent, but it depends on access to results that the search engine doesn’t offer through an official API, which adds continuity risk. That’s irrelevant for a weekend project and worth weighing for anything you plan to run for years.
### **What is the output shaped for?**
A free tier measured in requests tells you nothing about how many requests a question needs. APIs that return 20-word snippets built for people force your model to fetch each page separately, so one question can cost five calls. APIs that return dense, query-relevant excerpts often answer in one. Compare providers on cost per answer.
## **The best free web search APIs in 2026**
### **1. Parallel Search API**
The Parallel Search API[Parallel Search API] is built for AI agents rather than people reading a results page. It runs on Parallel’s own web index, which holds billions of pages and adds or updates millions daily[holds billions of pages and adds or updates millions daily], and returns ranked results with dense, query-relevant excerpts instead of short snippets. There are three ways to use it for free, and they stack:
- - **A free hosted MCP server.** Point any MCP client at mcp.parallel.ai/mcp[mcp.parallel.ai/mcp] and your agent gets web search and page fetch right away. It needs no account, API key, or card, and anonymous use is free at lower rate limits.
- - **A signup credit.** New accounts at platform.parallel.ai[platform.parallel.ai] can earn up to $80 in credits at signup[earn up to $80 in credits at signup], which you can spend across Parallel’s APIs.
- - **$5 in free credits every month.** Eligible organizations with a card on file get $5 in credits each month[$5 in credits each month], spent before any paid balance. Usage past the $5 is billed to that card at standard rates[Usage past the $5 is billed to that card at standard rates]. Unused credit expires at month end, it’s one credit per organization, and marketplace and postpaid organizations aren’t eligible.
The Search API has four modes[four modes]: Turbo (about 200ms, $1 per 1,000 requests), Fast (about 700ms, $1 per 1,000, and the mode the docs suggest starting with), Basic (about 1s, $5 per 1,000), and Advanced (about 3s, $5 per 1,000, and the default). Each request includes 10 results, and extra results cost $1 per 1,000[extra results cost $1 per 1,000]. Turbo supports English and Japanese only.
Parallel Web Systems is SOC 2 Type II certified[SOC 2 Type II certified], and zero data retention is available on Enterprise plans. The Search API quickstart[Search API quickstart] takes a few minutes end to end.
### **2. Tavily**
Tavily’s free Researcher plan gives 1,000 API credits a month with no credit card[1,000 API credits a month with no credit card]. A basic search costs 1 credit and an advanced search 2, so that’s 500 to 1,000 searches. Credits reset on the first of each month[Credits reset on the first of each month], and requests stop when they run out. Development keys are limited to 100 requests per minute[100 requests per minute]. Paid usage costs $0.008 per credit pay-as-you-go, and monthly plans bring that down to between $0.0075 and $0.005.
### **3. Exa**
Exa gives new accounts a $10 onboarding bonus plus $10 in credits every month, with no payment method required[$10 onboarding bonus plus $10 in credits every month, with no payment method required]. The free plan allows 10 search queries per second. Exa lists Search at $4 per 1,000 requests in Lite and $7 in Auto, so $10 covers about 1,400 to 2,500 searches a month.
### **4. TinyFish**
TinyFish makes its Search and Fetch APIs free[Search and Fetch APIs free], with no card required. Search is limited to 30 requests a minute and 500 an hour, and Fetch to 150 URLs a minute and 1,000 a day. Requests need an API key[API key], so you do sign up. On the Artificial Analysis Search Index[Artificial Analysis Search Index], TinyFish Search scores 71 at a measured search cost of $0. The rate limits are the catch for anything with fan-out.
### **5. Firecrawl**
Firecrawl’s free plan[free plan] gives 1,000 credits a month with no card and 2 concurrent requests. Search costs 2 credits per 10 results, so the plan covers about 500 searches, or 1,000 scraped pages. On the free plan, /search is limited to 10 requests a minute[/search is limited to 10 requests a minute]. Since June 2026 Firecrawl also offers Keyless[Keyless]: search and scrape with no account or key, capped per IP address per day. Firecrawl Search scores 73 on the Artificial Analysis Search Index[Artificial Analysis Search Index].
### **6. Serper**
Serper gives 2,500 free queries with no credit card[2,500 free queries with no credit card]. After that, credits come in prepaid packs that stay valid for six months. The $50 pack buys 50,000 queries ($1.00 per 1,000), and the rate falls to $0.30 per 1,000 only at the $3,750 pack of 12.5 million. What you get is Google results as JSON. The free grant is one-time, and like every Google SERP API, Serper depends on access to Google’s results that Google doesn’t offer through an official API, which adds continuity risk.
### **7. Linkup**
Linkup credits $20 when you sign up with a professional email and tops eligible accounts back up to $20 each month[credits $20 when you sign up with a professional email and tops eligible accounts back up to $20 each month]. Its pricing page frames that as 4,000 free queries[4,000 free queries], since standard search costs $5 per 1,000. SOC 2 Type II is included on all plans, and zero data retention is available on request[available on request]. The default rate limit is 10 queries per second per organization[10 queries per second per organization]. Linkup also published its SimpleQA evaluation script[its SimpleQA evaluation script], so you can rerun its accuracy claim yourself.
### **8. Brave Search API**
In February 2026 Brave replaced its free plan with $5 in monthly credits[replaced its free plan with $5 in monthly credits] on each plan. To claim the credit, you attribute Brave Search API on your project’s website or about page. Brave asks for a card at signup to prevent abuse[asks for a card at signup to prevent abuse] and sells prepaid credits, so you’re charged only for credits you choose to buy. Every result comes from Brave’s own independent index, with no fallback to another search engine. The Search plan costs $5 per 1,000 requests at up to 50 requests per second, so the credit covers about 1,000 searches a month.
### **9. SerpApi**
SerpApi’s free plan[free plan] gives 250 searches a month, with throughput of 50 searches an hour. It’s the smallest allowance here, enough to test the output format and a handful of real queries before you pay.
### **10. Jina Search**
Jina’s s.jina.ai endpoint[s.jina.ai endpoint] searches the web and returns the results as LLM-friendly text. It needs an API key, and every new key comes with 10 million free tokens. Each search costs a fixed number of tokens, starting from 10,000, so the grant covers up to about 1,000 searches, at 100 requests a minute on a free key.
### **11. Google Custom Search JSON API**
The Custom Search JSON API[Custom Search JSON API] was the historical default: 100 queries a day free, $5 per 1,000 after that, up to 10,000 a day. It’s closed to new customers, and existing customers have until January 1, 2027 to move. Google calls Vertex AI Search “a favorable alternative for searching up to 50 domains” and asks teams that need full web search to contact it. If you don’t already have a key, you can’t use this one.
### **12. Self-hosted SearXNG**
SearXNG[SearXNG] is a self-hosted metasearch engine with no per-query cost. If “free” means “no invoice,” it’s the only entry that fully qualifies. You run the server and own the reliability. Results come from public engines, so you inherit their rate limiting and blocking, and there’s no SLA or support. JSON output has to be enabled in settings.yml, or the API returns a 403, and many public instances leave it off. It suits privacy-sensitive internal tools and experiments that can tolerate failure.
## **Is there a free Google search API?**
Not for new projects. Google’s Custom Search JSON API[Custom Search JSON API] is closed to new customers and ends for existing ones on January 1, 2027. The free routes to Google results are third-party SERP APIs: Serper’s one-time 2,500 queries and SerpApi’s 250 a month. Microsoft retired the Bing Search APIs on August 11, 2025[retired the Bing Search APIs on August 11, 2025] and points developers to Grounding with Bing Search in Azure AI Agents. If you need web results for an AI agent rather than Google’s ranking specifically, free tiers including those from Parallel, Tavily, Exa, TinyFish, and Firecrawl return content shaped for a model.
## **Free tiers side by side**
Checked against each vendor’s pricing and documentation pages on October 27, 2026.
| Provider | What’s free | Card required | Free rate limit | Refills |
|---|---|---|---|---|
| Parallel | Hosted MCP server (no account), up to $80 to earn at signup, $5/month | Only for the monthly credit | Lower limits for anonymous MCP use. API: 600 Search requests/min (default for all accounts) | Monthly |
| Tavily | 1,000 credits/month | No | 100 requests/min (development key) | Monthly |
| Exa | $10 at signup, then $10/month | No | 10 queries/sec | Monthly |
| TinyFish | Search and Fetch | No (API key required) | Search: 30/min, 500/hour | Ongoing |
| Firecrawl | 1,000 credits/month, or Keyless with no account | No | 2 concurrent, /search 10/min | Monthly (Keyless daily) |
| Serper | 2,500 queries | No | Not stated for free queries (paid packs from 50/sec) | One-time |
| Linkup | $20 at signup, topped up to $20/month (about 4,000 searches) | No | 10 queries/sec | Monthly |
| Brave | $5/month credit with attribution (about 1,000 searches) | Yes, to prevent abuse (prepaid credits only) | 50 requests/sec | Monthly |
| SerpApi | 250 searches/month | Not stated | 50 searches/hour | Monthly |
| Jina (s.jina.ai) | 10 million tokens per new key (about 1,000 searches) | Not stated | 100 requests/min | One-time |
| Google Custom Search | 100 queries/day, closed to new customers, ends January 1, 2027 | For overage | 10,000 queries/day | Daily |
| SearXNG | Unlimited, self-hosted | No | Set by you and upstream engines | N/A |
Check official documentation for current pricing.
## **How do you add free web search to an agent without an API key?**
Two options on this list work with no account at all: Firecrawl Keyless and the Parallel MCP server[Parallel MCP server]. The Parallel server is a hosted, streamable-HTTP MCP server at https://mcp.parallel.ai/mcp. Without a key it exposes two tools, free at lower rate limits:
- - **web_search** searches the web and returns excerpts that usually hold enough to answer the question. It runs the Search API in Fast mode by default.
- - **web_fetch** reads specific URLs and returns the relevant parts of each page, or the full page as markdown.
A third tool, web_research, answers a single question with citations in one call. It needs you to sign in or use an API key. For Claude Code or Codex, adding the server is one command:
1claude mcp add --transport http parallel https://mcp.parallel.ai/mcp```claude mcp add --transport http parallel https://mcp.parallel.ai/mcp```
1codex mcp add parallel --url https://mcp.parallel.ai/mcp```codex mcp add parallel --url https://mcp.parallel.ai/mcp```
For clients configured through JSON, the entry is a few lines:
12345678{
"mcpServers": {
"parallel": {
"type": "http",
"url": "https://mcp.parallel.ai/mcp"
}
}
}``` { "mcpServers": { "parallel": { "type": "http", "url": "https://mcp.parallel.ai/mcp" } }}``` parallel.ai/mcp[parallel.ai/mcp] has one-click installs for Claude, ChatGPT, Cursor, and VS Code, and Cursor users can also add it from the Cursor Marketplace[Cursor Marketplace]. The MCP quickstart[MCP quickstart] covers other clients, including Gemini CLI, Windsurf, and Zed.
For higher rate limits, create an account at platform.parallel.ai[platform.parallel.ai] and send your API key as a Bearer token, or connect to https://mcp.parallel.ai/mcp?login to sign in over OAuth. Signed-in calls are billed at the same rates as the underlying API[billed at the same rates as the underlying API], so they draw on your signup and monthly credits. For production traffic, call the Search API directly, where you choose the mode, the result count, and the cost per request.
## **How do Parallel’s four search modes change the free-tier math?**
Parallel sells four search modes at two prices, so spend can match what a query is worth.
| Mode | Latency | Price | Built for |
|---|---|---|---|
| Turbo | ~200ms | $1 / 1,000 requests | Chat, voice, RAG pre-filtering, high-volume lookups |
| Fast | ~700ms | $1 / 1,000 requests | Most agent loops (the docs’ recommended starting mode) |
| Basic | ~1s | $5 / 1,000 requests | Quick retrieval with longer excerpts per result |
| Advanced | ~3s | $5 / 1,000 requests | Highest-quality, multi-hop results (the default) |
If price brought you here, look at Turbo and Fast. On parallel.ai/benchmarks[parallel.ai/benchmarks] (evals run September 9, 2026), with a GPT-5.6 Luna agent at the low-cost tier, Fast scored 94% on a 100-question sample of SimpleQA Verified and Turbo 91%, both at a total cost of $2 per 1,000 questions, including model tokens and tool calls, against 94% at $5.50 for Perplexity and 94% at $17.40 for Tavily. Those are Parallel’s own runs.
The independent Artificial Analysis Search Index[Artificial Analysis Search Index] (latest data October 19, 2026) runs the same GPT-5.6 Luna answer model against every provider and changes only the search API. Parallel Search (advanced) scores 75, level with Brave (LLM context) and behind Exa (auto, dynamic highlights) and Perplexity Search (medium) at 80, Exa (lite) and Perplexity (high) at 79, and Octen and Perplexity (low) at 77. Fast scores 73 at $8.41 of search cost per 1,000 benchmark tasks, the lowest search cost of any paid provider measured, and Turbo scores 67 at $13.64. TinyFish, which is free, measures $0. Counting the answer model’s tokens as well, Fast comes to about $27.37 per 1,000 tasks, and Octen, which scores 77, costs less at about $23.99. Firecrawl Search also scores 73.
Each provider’s row on that index carries its own measurement date, and Parallel’s rows were measured in August 2026. Run your own evaluation on your own queries before you commit.
## **How far does $5 a month go on Parallel?**
Parallel prices per request, so the monthly credit turns into a number you can plan against. At current prices[current prices], $5 covers up to:
- - 5,000 searches in Turbo or Fast mode, or 1,000 in Basic or Advanced
- - 5,000 Extract API requests
- - 1,000 Task API runs on the lite processor
- - 1,666 Monitor API executions on the lite processor
Or any mix. 5,000 searches a month is about 165 a day, enough for a personal research agent, a Slack bot, or a monitoring job. You pay only for successful Task runs, so a broken query doesn’t eat the allowance. Free credit is spent before paid balance, so adding a card doesn’t make you pay sooner. It decides what happens once the free credit is gone. Usage past the $5 is billed to the card at standard rates, and Parallel’s monthly spend limits email admins but don’t block requests.
## **Which free option fits your project?**
- - **You want web search in your coding agent in the next minute.** Parallel’s hosted MCP server, or Firecrawl Keyless. No key needed.
- - **You’re prototyping and want to avoid an invoice.** Tavily, Exa, or TinyFish. None of them needs a card.
- - **You need Google’s ranking.** Serper, with 2,500 free queries and $1.00 per 1,000 on its $50 pack. Accept the dependency on Google’s results knowingly.
- - **The free tier is a trial for something headed to production.** Evaluate the paid path: the mode you’ll ship, the rate limits you’ll need, and whether SOC 2 and zero data retention are included or an enterprise add-on.
- - **Volume is high and per-query cost decides the build.** Parallel Turbo or Fast at $1 per 1,000 requests. Measure cost per answer rather than cost per request.
- - **You can’t spend money and can run infrastructure.** Self-hosted SearXNG, with clear eyes about reliability.
## **Frequently asked questions about free web search APIs**
### **Is there a web search API that’s free with no credit card and no signup?**
Two. The Parallel MCP server at https://mcp.parallel.ai/mcp needs neither, and neither does Firecrawl Keyless, which is capped per IP address per day. Tavily, Exa, TinyFish, Linkup (with a professional email), Firecrawl’s free plan, and Serper need an account but no card. Brave asks for a card to prevent abuse and charges only for prepaid credits you choose to buy. Parallel’s $5 monthly credit needs a card on file.
### **Can I run a production application on a free tier?**
Rate limits usually rule it out before credit caps do. Free tiers are throttled, with low concurrency and no SLA, because they’re built for evaluation and hobby use. Monthly credits can carry a low-volume production job, such as a monitoring task or an internal bot, but anything user-facing will hit the rate limit first.
### **What happened to the free Bing and Google search APIs?**
Microsoft retired the Bing Search APIs on August 11, 2025[retired the Bing Search APIs on August 11, 2025] and points developers to Grounding with Bing Search in Azure AI Agents. Google’s Custom Search JSON API[Custom Search JSON API] is closed to new customers and ends on January 1, 2027. Both historical free defaults for web search are gone.
### **Do free credits roll over if I don’t use them?**
Mostly not. Parallel’s $5 monthly credit expires at month end, and Tavily’s credits reset on the first of the month. Linkup tops your balance back up to $20 rather than adding $20 on top. Serper’s paid credits stay valid for six months. Assume no rollover unless a provider says otherwise in writing.
### **What’s the cheapest paid web search API once the free tier runs out?**
By list price per request, Parallel Turbo and Fast at $1 per 1,000 are the lowest paid rate in this guide among APIs that run their own index. Serper starts at $1.00 per 1,000 and reaches $0.30 only on its $3,750 pack. Cost per task is the better measure. The Artificial Analysis Search Index[Artificial Analysis Search Index] (latest data October 19, 2026) counts search cost plus the answer model’s tokens. On that basis TinyFish’s free tier costs about $9.30 per 1,000 tasks, Octen about $23.99, Parallel Turbo about $26.06, and Parallel Fast about $27.37. Serper isn’t on that index.
### **How do I move from the free MCP server to the API?**
Create an account, generate a key, and call the Search endpoint directly. The quickstart[quickstart] covers it in a few minutes. You keep the same index and excerpt format, and gain mode selection, result-count control, and higher rate limits. Full rates are on the pricing page[pricing page].
**Related reading: **The honest 2026 comparison: web search APIs for AI agents[The honest 2026 comparison: web search APIs for AI agents] / Tavily vs. Parallel[Tavily vs. Parallel] / Exa vs. Parallel[Exa vs. Parallel] / Brave Search API vs. Parallel[Brave Search API vs. Parallel]