Braintrust pricing tells you exactly what scoring your agent costs. It never tells your customer what the agent was worth.

Braintrust's pricing runs three tiers, and every number below comes from Braintrust's own current pricing page rather than a list price a third party writeup forgot to update. Starter costs $0 a month, includes ten dollars of model credit, covers one gigabyte of processed data and ten thousand scores, holds fourteen days of retention, and still ships with unlimited users, projects, datasets, playgrounds, and experiments, no seat cap even on the free tier. Pro costs $249 a month flat, no per seat charge as a workspace adds engineers, raises the included allowance to five gigabytes of data and fifty thousand scores, stretches retention to thirty days with paid extended storage past that, and adds custom charts, environments, priority support, basic role based access control, a loop agent for automated evaluation runs, playground annotations, and unlimited human review scores.

Enterprise is where Braintrust's own pricing page stops publishing numbers entirely, listing the tier only as custom pricing, quoted through sales, with no minimum contract value or usage based starting point shown anywhere on the page. What the page does document is the feature set added at that tier: full role based access control instead of Pro's basic version, custom retention rules set per project rather than one account wide window, S3 export for a company's own long term storage, on premises or hosted deployment for teams that cannot send evaluation data to a third party, a guaranteed SLA, and a signed BAA for HIPAA covered workloads. Several third party pricing writeups list a specific Enterprise dollar figure anyway; none of those numbers appear on Braintrust's own site, so treat any exact figure you see elsewhere as unverified until Braintrust's own sales team quotes one against real usage.

Every tier above bills against the same two meters, processed data and scores, and Braintrust's own documentation defines a score precisely: it is any scored output an evaluation run produces, whether an LLM acts as the judge, an autoeval built into the platform runs it, or a team's own custom code scorer grades it. A human reviewer manually grading a trace during review generates a score the same way, which is why Pro lists unlimited human review scores as its own named feature rather than folding that activity into the fifty thousand score allowance. Running one autoeval and one custom scorer across the same dataset produces two scores per row, not one, the detail that most often explains why score usage climbs faster than a team's row count would suggest on its own.

Past the included allowance, Braintrust does not cut a workspace off; it keeps ingesting and bills the overage per unit rather than in blocks. On Starter, data past the first gigabyte costs $4 per additional gigabyte and scores past the first ten thousand cost $2.50 per thousand. On Pro, the same two meters drop to $3 per additional gigabyte and $1.50 per thousand scores, plus fifty cents per gigabyte per month for data retained past the included thirty days. Take a Pro workspace processing eight gigabytes of data and running eighty thousand scores in a month: the base $249 covers the first five gigabytes and first fifty thousand scores, the remaining three gigabytes add $9, and the remaining thirty thousand scores add $45, landing the real platform bill at $303, before counting whatever model credit the team burns past the included $249 running evaluations through Braintrust's own proxy.