Braintrust Pricing in 2026: What You Actually Pay for LLM Evals

Quick answer (2026): Braintrust has three tiers. Starter is free, Pro is $249 a month, and Enterprise is custom. The sticker price is almost never your real bill. Braintrust meters four separate things, and for most teams one of them, the evaluation "scores" meter, is what the invoice is actually made of. Starter includes 10,000 scores a month and charges $2.50 per 1,000 after that. Pro includes 50,000 scores and charges $1.50 per 1,000 after that. That single meter is why two teams running what looks like "the same" eval suite can be billed $5 and $3,378 in the same month. Everything below is sourced from Braintrust's pricing page (2026).

The plans, decoded (2026)

Two things the table hides. First, both tiers also hand you model credits ($10 on Starter, $249 on Pro) that only apply to Braintrust's built-in models and Topics. If you bring your own API key, those credits are moot, and on Pro the $249 sticker becomes a pure platform fee that you recoup only through the cheaper overage rates and the larger free blocks. Second, and this is the important one, both tiers include unlimited users, projects, datasets, playgrounds, and experiments. Your headcount does not move this bill. Usage does.

What Braintrust actually meters: "scores"

A score is one output from a scorer logged against one row of an eval. A scorer can be an LLM-as-a-judge call, a built-in autoeval, or your own code check. The meter is not per eval run, and it is not per seat. It is:

scores = rows x scorers x runs.

That multiplication is where the surprises live. Add a fourth scorer to a suite and every run costs a third more. Move the suite from "run it when I remember" to "run it on every pull request" and the run count becomes your CI traffic, not your team size. A 500-row suite with 6 scorers is 3,000 scores per run before you have measured anything twice. The seat count never enters the equation.

The number that decides your plan: about 199,000 scores a month

Ignore data for a second and price the scores meter on its own. On Starter you pay $2.50 per 1,000 scores above the free 10,000. On Pro you pay $249 flat plus $1.50 per 1,000 above the free 50,000. Set those two lines equal and they cross at roughly 199,000 scores a month.

Below about 199,000 scores a month, Starter plus overage is cheaper than Pro. Above it, Pro's lower per-score rate and larger free block win, and the $249 stops being a sticker and starts being a discount. Most teams never do this subtraction. They either default to Pro far too early or cling to Starter well past the point where the overage rate is quietly costing them more than the subscription would.

Three 30-day bills, same product

Team A runs a small suite a few times a week. At 12,000 scores it is 2,000 over the free block, so $5.00, and under 1 GB of data means no data charge. Braintrust's free tier genuinely covers real solo eval work.

Team B is the trap. Six scorers, a 500-row regression suite, running in CI on every pull request plus nightly, is about 1,350,000 scores a month. On Starter that is (1,350,000 - 10,000) / 1,000 x $2.50 = $3,350 in scores plus $28 of data overage, or about $3,378. On Pro it is $249 plus (1,350,000 - 50,000) / 1,000 x $1.50 = $1,950 in scores plus $9 of data, about $2,208. The $249 sticker is a rounding error next to the meter. And the driver was scorer sprawl on the CI path, not the size of the team.

Team C flips the story. Only two scorers, so just 60,000 scores, but 60 GB of production traces logged. Here the bill is data, not scores. On Starter it is $125 of scores plus $236 of data, about $361. On Pro it is $264 of scores plus $165 of data, about $429. Because Team C sits below the ~199,000-score crossover, Starter is actually the cheaper plan even though it logs a lot. Pro is not automatically the "serious team" choice.

The one lever that cut Team B's bill 71%

You do not need all six scorers on every pull request. Split the suite by path:

• Nightly, full coverage: all 6 scorers on all 500 rows, 30 runs a month = 90,000 scores.
• Per pull request, smoke signal: 1 scorer on all 500 rows, ~420 runs = 210,000 scores.

Total drops from 1,350,000 to 300,000 scores a month. Team B's Pro bill falls from about $2,208 to about $633, a 71% cut, with the full six-scorer regression still running every single night. Same tests, same rows, cheaper signal on the noisy CI path. If you also sample rows on pull requests, say 100 of 500, it drops further. The cheapest fix here was never a plan change. It was running fewer scorers where the runs are frequent.

Which meter is your driver?

Before you pick a plan, tag your own driver with two quick numbers. Multiply rows x scorers x runs for your scores figure, and estimate GB logged per month for your data figure. Whichever is bigger tells you which lever to pull, and, as Team C showed, can flip which plan is cheaper. Scores are the headline for teams doing heavy evaluation; processed data is the quiet one for teams logging large production traces. Do not assume you know which camp you are in until you have done the multiplication.

The escape hatch: run the scorers yourself

The scorers are open source. Braintrust maintains autoevals under the MIT license, the same LLM-as-a-judge, heuristic, and statistical scorers the hosted product runs. Run them in your own CI and log pass or fail to a store you own, and the scores meter drops to zero. You pay compute you already have instead of $1.50 per 1,000. What you give up is the hosted experiment UI, regression diffing, and shared experiments, which is precisely what you are renting on Pro.

That is the honest line. Managed wins when the diffing UI and shared experiments are real product surface for your team. Roll-your-own wins when all you need is a red-or-green gate in CI. If you want a middle ground, hosted tracing without an eval-scores meter, an open-core tool like Langfuse prices tracing differently (see our Langfuse pricing teardown), and the same build-versus-buy crossover applies.

Math check

A score is rows times scorers times runs, and Braintrust charges by the score, not the seat. Count that product before you count anything else. Below about 199,000 scores a month, stay on Starter and pay overage. Above it, Pro pays for itself. And if the number is scaring you, the cheapest move is usually not a plan change. Math check: 1,350,000 scores becomes 300,000 with one nightly-versus-per-PR split, and a $2,208 bill becomes $633, same nightly coverage.