Braintrust

What Braintrust actually meters

Two meters run: bytes in and scores recorded. Processed data counts the bytes you send across logs, experiments and datasets: inputs, outputs, prompts, metadata, traces and spans, dataset rows and attachments. Braintrust measures it at ingestion, so a retention policy or a deletion changes nothing about that month's invoice. The only way down is to send less.

Scores count one by one, and experiment scores, online scoring rules, human reviews and custom code scorers all increment the same counter. Scoring a span costs extra on both meters at once: the score bills, and each scorer run lands in the trace as a score span, billing again as processed data.

An LLM-as-a-judge adds a third charge when the judge is a built-in model. Those tokens draw on your model credits, then bill from $0.15 per million input tokens for GLM-5.3 Flash to $50.00 per million output for GPT-6 Astra. Use your own key and your provider bills you instead. Jev, Braintrust's own model, is free for scorers.

Per span or per trace: the arithmetic

LangSmith charges $0.005 per trace and lets one trace hold 25,000 runs, so depth is free. Braintrust prices that depth by weight. Assume a 40-span run averaging 8 KB of ingested payload per span, at 100,000 runs a month: 0.32 MB a run, 32 GB a month.

On Pro, 32 GB less 5 GB included leaves 27 GB at $3, so $81, plus the $249 fee: $330. LangSmith Plus bills 100,000 traces less 10,000 included at $0.005, so $450, plus a $39 seat: $489. Braintrust wins until you score. Add two scores per run and 150,000 of those 200,000 scores sit past the Pro allowance at $1.50 per 1,000, which is $225, taking Braintrust to $555. Span size and score count are assumptions, and both are the levers.

How model credits work

Model credits are a monthly allowance for Braintrust-served inference, not for tracing. They cover the built-in models and the Topics pipeline, and they're worth $10 on Starter, $100 on Pro and a custom figure on Enterprise.

They don't roll over, and Braintrust says so on the pricing page, in the plans doc and in the billing docs, which beats most vendors. No top-up pack, pooling rule or deduction order is documented; both uses draw on one balance. At zero, Topics bills $0.06 per million input tokens and $0.40 per million output. On Starter with no card, Topics and built-in models pause until the next cycle.

What happens when you hit the limit

Starter blocks. Pro and on-demand auto-charge. Exceed the Starter allowances with no card on file and usage above them stops, which makes Starter a hard cap rather than a soft one. Add a card and on-demand switches on: Braintrust charges $1 to validate it, then bills everything above the allowance on the 1st. Disabling it blocks again immediately.

Pro has no documented spend cap. The closest control is the sampling rate on an online scoring rule. A failed payment gets a grace period with access intact, then a downgrade to Starter.