CrewAI Pricing: What It Actually Costs

Costing a multi-agent deployment is annoying in a specific way. The framework is free, so finance asks why there is a line item at all, and by the time you have a real answer you are three vendors deep. CrewAI sits at the center of that question for a lot of US engineering teams right now, and the public information about what it costs is worse than it looks.

Here is what the vendor itself publishes, read firsthand today rather than copied from a software directory, plus the two bills that arrive alongside it and usually dwarf it.

CrewAI pricing plans in 2026

Two rows. That is the entire published price list.

The open source framework is separate from both and stays free. You can build and run crews on your own machines forever without touching either plan. What the paid product sells is the platform around them: deployment, monitoring, governance and a managed place for crews to run, which CrewAI calls AMP, its agent management platform.

Why most CrewAI pricing articles show a tier that is not on the page

Search for this topic and you will find a consistent story across half a dozen directories and blogs: Basic free, Professional at $25 a month with 100 executions and $0.50 per execution after that, then Enterprise. Some go further and put a dollar range on Enterprise, in the tens of thousands a year.

None of that appears on the vendor's pricing page as it stands today. The Professional tier is absent, the overage rate is absent, and CrewAI publishes no Enterprise figure at all, so any specific annual number you read is somebody's estimate presented as a fact. It may have been accurate once. Agent platforms have repriced repeatedly over the last eighteen months and directory pages are not maintained after publication, which is exactly why the standing advice on this site is to read the vendor's own page before you budget anything.

The practical consequence for a buyer is small but real: do not build a business case on a $25 line item you cannot actually purchase. If you need something between free and a sales call, that is a question for the sales call.

What counts as a workflow execution, and why 50 goes quickly

An execution is one run of a deployed crew, not one agent action and not one model call. That sounds generous until you attach it to anything scheduled. A crew that runs hourly against a watchlist burns 50 executions in about two days. A crew triggered by an inbound webhook burns them at whatever rate your traffic dictates. Fifty a month is an evaluation allowance, and it is honest of CrewAI to frame Basic that way rather than pretending it is a production tier.

The number that does not appear anywhere on the price list is far more important: what each execution costs you downstream. That is a function of how many model calls the crew makes and how much text is in each one, and it is entirely under your control.

The three bills a CrewAI deployment actually generates

The second row is where the surprises live, and the mechanism is worth spelling out because it is not obvious from the CrewAI docs. Tool output goes into the message history. It stays there for the rest of the task. Every subsequent model call in that task re-sends it as input tokens. So a tool that returns a large blob is not charged once, it is charged once per remaining step.

That interacts badly with the default scraper. CrewAI's built-in ScrapeWebsiteTool fetches a page with requests, parses it with BeautifulSoup and returns the full page text. Navigation, cookie notices, sidebars, related posts and footers all come along. On a typical marketing or documentation page the useful content is a minority of the words, and you pay for the rest repeatedly. Returning clean markdown instead is the single cheapest optimization available to a crew that reads the web, and it usually cuts more spend than switching to a smaller model would.

What the web data slot costs

CrewAI documents fourteen tools in its web scraping category and eleven of them require an API key from another vendor, so this line item is not optional for most production crews. The full table of which tool renders JavaScript, which can crawl, and which needs a paid account is on our CrewAI web scraping tool page, along with a working BaseTool you can paste into a crew.

Rates vary more than people expect because vendors meter different verbs. Credit-based APIs charge a multiplier when a page needs a browser, often five or six times the plain rate, which means a headline price per thousand credits tells you very little until you know what share of your pages are JavaScript heavy. Runtime-based platforms charge for how long a scraper runs rather than how many pages it returned. ClawEngine is flat per page with rendering included, from $39 a month, which is the model we prefer precisely because it survives contact with a budget spreadsheet. If you are comparing across the field, the per thousand page arithmetic for every vendor CrewAI lists is worked out on our best web scraping API comparison, and the credit multiplier detail specifically is on the Firecrawl alternative page.

Is CrewAI free?

The framework is, permanently, and it is a real open source project rather than a trial. The platform has a free Basic tier with 50 workflow executions a month. What is not free is running crews in production with governance, monitoring and a deployment target, which is what Enterprise sells, and it is not free to run the models or fetch the pages either. A fair summary is that CrewAI is free to build with and priced on request to operate.

How much does CrewAI Enterprise cost?

CrewAI does not publish a figure and asks you to request a trial. Treat any specific annual number you find in a directory as an unverified estimate. What the page does tell you is what shapes the quote: seats, deployment model (CrewAI cloud, your own VPC, or your own infrastructure), the governance feature set, and the 45 day onboarding. Go into the call with your execution volume, your deployment preference and your compliance requirements written down, because those are the variables the number is built from.

The onboarding length is itself a useful signal. A 45 day program plus optional forward deployed engineering says that a first production crew is not a weekend project, and teams without agent engineering experience in house frequently end up paying for that expertise somewhere, whether through the vendor or by bringing in AI developers on contract. Budget for it rather than discovering it in month two.

How to cost your own crew in an afternoon

Do not model this from published rates. Measure it, on the crew you actually built.

Run one representative execution with verbose logging on. Count the model calls it makes and record the input token count for each. Sum them and multiply by your model's input rate, then multiply by executions per month. That is bill two, and for most crews it lands several times larger than any platform fee.

Then run the same execution twice more, once with your scraping tool returning raw page text and once returning clean markdown. Compare the token totals. The gap between those two numbers, multiplied by your monthly execution count, is the actual value of the web data vendor decision, and it is almost always larger than the difference between the vendors' own price lists.

Finally, count the distinct URLs fetched per execution and check for duplicates. Agents in the same crew do not share memory of what a sibling fetched, so a researcher, an analyst and a writer will each pull the same page. Caching in front of the tool is free and typically removes a third of the fetches on research-shaped work.

The recommendation

If you are evaluating, start on Basic and accept that 50 executions is a demo budget. If you are deploying, get the Enterprise quote early, because the deployment model question (their cloud, your VPC, your metal) changes the answer materially and it is not something you want to discover after the architecture is set.

Whichever way that goes, spend an afternoon on the token arithmetic before you spend a week on vendor selection. The platform fee is a known quantity you can negotiate once. The tokens your agents burn reading badly formatted web pages are a recurring cost you control entirely, and the fix is a one line change to what your scraping tool returns.