How Much Are Free Search API Credits Worth?
Search API credits come in two shapes: a standing free tier of a few hundred to a few thousand queries a month that any signup gets, and larger startup grants from providers running credit programs. The free tiers alone usually carry you through building and evaluating before you pay anything.
That makes search unusual among AI infrastructure bills. You do not need a grant to start. You need one the week your agent starts issuing six searches per question instead of one.
AI Perks tracks which providers in this category currently run credit programs, alongside $7.7M in credits from 194 companies.
| Provider | What it queries | Billed by | Free allowance shape | Best fit |
|---|---|---|---|---|
| Exa | Own neural index of the web | Search call, plus content | Signup credits | Semantic "find pages like this" retrieval |
| Tavily | Aggregated results, cleaned for models | Search call | Standing monthly allowance | Agent tool calls, minimal glue code |
| Brave Search API | Independent crawler and index | Query | Monthly allowance, rate limited | Keyword search, no Google dependency |
| Firecrawl | Extracts pages as clean markdown | Page or credit | One-time signup allowance | Known URLs into model-ready text |
| Serper | Google results via a proxy | Query | Signup credits | Cheap high-volume SERP data |
| SerpAPI | Google and other engines via proxies | Search | Small monthly allowance | Structured SERP fields across many engines |
| Perplexity Sonar | Search plus a grounded answer | Tokens plus search | Allowance bundled with some plans | Search and synthesis in one call |
| Google Programmable Search | Restricted slice of Google | Query | Small daily allowance | Site-scoped or domain-limited search |
Free tier figures move often here. Read the shape of each allowance rather than the number, and re-check terms before building a cost model on them.
What a Search API Actually Does for an AI Product
A search API is how a language model learns about anything after its training cutoff, or anything on a page it never memorised. It is the live half of retrieval, where a vector database is the private half.
Three jobs cover most real usage. Grounding puts fresh pages in front of the model before it answers, the difference between a confident wrong answer and a cited correct one. Research agents search, read, work out what they still do not know, and search again, which is why search volume grows faster than user count. Extraction pipelines for enrichment, price monitoring and lead research use search as discovery and treat clean content as the product.
Four different products get called "search API" and they bill differently. SERP proxies such as Serper and SerpAPI fetch Google's results for you. Independent indexes such as Brave run their own crawler. Neural indexes such as Exa retrieve by meaning rather than keyword. Extraction services such as Firecrawl rank nothing, they turn a page into clean text. Most production systems use two of the four.
How Search API Costs Behave at Scale
Search billing is per query and close to linear, which makes it the most predictable AI bill you have and the easiest to multiply by accident. Published rates across the category run roughly $1 to $15 per thousand queries depending on provider and volume tier.
Linear is comforting until you count the multiplier. One user question rarely means one search: query rewriting fans out into three to six searches, results trigger five to twenty page reads, and a follow-up round adds more. A single "research this company" turn can be thirty upstream calls, which at list prices is cents, not dollars. That is exactly why teams stop watching right before the bill matters.
Here is the part most cost models miss. The search call is usually the cheaper half of a search. Ten results at roughly 1,500 tokens of extracted text each is about 15,000 input tokens fed to a model. Against a model in the $3 per million input token band, that costs more in model tokens than in search fees, often by five times or more.
The practical consequence: reranking and truncating results before they reach the model saves more money than switching search providers. Fund both layers anyway, since AI Perks covers model and search credits in one place.
How to Choose Between Exa, Tavily, Brave, Firecrawl and SERP APIs
Pick by what you need back, not by benchmark scores. Ranked links mean a SERP proxy or an independent index. Meaning-based discovery means a neural index. Clean text from pages you already know about means an extraction service.
A short decision guide:
You want an agent tool working this afternoon. Tavily and Exa return model-ready results with almost no glue code. Prototype speed matters more than per-query price at your volume.
You need Google's actual ranking. Only SERP proxies give you that. You are paying someone to fetch results from an engine that would rather they did not, so treat it as a dependency with continuity risk.
You want to own your search dependency. Brave runs an independent index, which removes the proxy layer entirely. Microsoft's retirement of its Bing Search APIs in 2025 pushed a large cohort of teams onto independent providers, a supply shock nobody had modelled.
Your query is "find me things like this", not "find me this". Neural retrieval is better at that shape, keyword search is better at product codes, error strings and proper nouns. Many teams run both.
You mostly have URLs already. Then you need extraction, not search. Firecrawl and similar services are a different line item and usually a smaller one.
Which of these run credit programs changes quarter to quarter, and current status for each is listed at getaiperks.com.
What Order to Apply In
Exhaust the standing free tiers first, since they need no application and no approval. Apply to startup programs only for the provider you have decided to ship on, then to the adjacent layers whose credits stack.
First, the free tiers. Sign up for three or four providers and run the same hundred real queries through each. This costs nothing and settles the choice with evidence rather than a comparison table.
Second, the provider you chose. A grant is worth applying for once you know it funds infrastructure you will still use in a year. Requirements differ for every program above, and current status for each is tracked at getaiperks.com.
Third, the layers underneath. Search credits are among the smaller grants in the market, and the bill they offset is smaller than your model or compute bill. Funding only search funds the cheapest layer.
Last, anything time-boxed. A grant that starts counting down before you have traffic burns most of its value on an idle integration. Trigger it the month you ship.
One nuance worth checking: some third-party APIs can be bought through a cloud marketplace and drawn against committed cloud spend, but startup credit grants frequently exclude marketplace purchases. Confirm rather than assume.
What Founders Get Wrong About Search API Credits
The most common mistake is optimising the search bill while the model bill triggered by search results runs five times larger. The second is forgetting that rate limits, not credits, break a free tier.
Four patterns worth avoiding:
Ignoring rate limits. A free tier rated at one query per second will not serve an agent that fans out six parallel searches. The allowance is not the constraint, concurrency is.
Never caching. Consumer products see enormous query overlap between users. A cache keyed on the normalised query, with a short time to live, removes a large share of upstream calls at zero quality cost. Few teams build it before the bill hurts.
Treating search and retrieval as one decision. A search API reads the public web. A vector database reads your documents. Products that need both and fund only one rebuild the missing half under deadline.
Building on a single provider with no abstraction. Search providers change pricing, get acquired and retire APIs. A thin interface over your search call costs an afternoon and has repeatedly saved teams a quarter.
Founders who fund the whole stack rather than one layer end up furthest ahead. AI Perks tracks $7.7M across 194 companies, so search calls, model calls and compute can all be covered at once.
Frequently Asked Questions
How do I get free search API credits for an AI agent?
Start with the standing free tiers, which most providers in this category offer on signup with no application at all. Larger grants come through startup credit programs, where requirements differ by provider. Which providers currently run one is tracked at getaiperks.com.
Which search API is cheapest for high volume?
SERP proxies are generally cheapest per query at volume, with published rates in the category running roughly $1 to $15 per thousand queries depending on provider and tier. Cheapest per query is not cheapest overall, since result quality determines how many queries your agent needs per question.
Do I still need a vector database if I have a search API?
Usually yes, because they cover different corpora. A search API reads the public web, while a vector database searches your own documents, tickets and product data. Products that answer questions about both need both layers, and each has its own credit programs listed at getaiperks.com.
Can I stack search API credits with OpenAI or Anthropic credits?
Yes, and you should. They are separate bills from separate vendors. Search credits pay for finding and fetching pages, while model credits pay for reading and reasoning over them, which is typically the larger of the two. Holding both is how teams cover a full first year.
What does a search API cost per month in production?
A modest product running a few thousand searches a month typically spends tens of dollars. A research agent with heavy query fan-out can reach hundreds or low thousands at the same user count, because volume tracks agent steps rather than users. Model up from calls per task, not from signups.
Is scraping Google through a SERP API risky?
It carries real continuity risk rather than immediate legal risk for you as a buyer, since your provider is the one fetching the results. Engines change their markup and their tolerance, and retirements happen with little notice. Keeping a second provider integrated behind a thin interface is cheap insurance.
Search the whole web. Let someone else pay for the queries.