Braintrust is an evaluation-first platform: experiments, datasets, scorers and a playground, with logging and tracing of production traffic attached. Its pricing is unusual in the category because it does not count spans, traces or seats. It meters bytes. This post explains the plans and the two meters, and then works three monthly bills, which needs an assumption about how big a span is, so that assumption is stated.
Prices are as published on the Braintrust pricing page in October 2026.
The three plans
| Starter | Pro | Enterprise | |
|---|---|---|---|
| Price | $0 a month | $249 a month | Custom |
| Processed data | 1 GB a month, then $4 per GB | 5 GB a month, then $3 per GB | Custom |
| Scores | 10,000 a month, then $2.50 per 1,000 | 50,000 a month, then $1.50 per 1,000 | Custom |
| Model credits | $10, then token rates | $100, then token rates | Custom |
| Retention | 14 days | 30 days, then $0.50 per GB per month | Custom |
| Users, projects, experiments | Unlimited | Unlimited | Unlimited |
| Support | Community | Priority | Premium, SLAs |
Enterprise adds SAML SSO, custom retention and export to S3, a DPA and BAA, an uptime SLA and on-prem or dedicated hosting. Pro offers 6 to 12 months free to qualifying startups.
The two meters
Processed data is the bytes Braintrust ingests: inputs, outputs, metadata, span payloads. There is no span count. A thousand short chatbot turns and ten long-context agent runs can be the same number of gigabytes. What moves the meter is large prompts, long outputs and big tool results, which is what agents produce.
Scores are the results of scorers: LLM-as-judge evaluations, code-based checks, human review. Every score on every logged span counts. Scoring 100% of production traffic with three scorers is three scores per span.
Three worked months
Braintrust does not publish a bytes-per-span figure, so an assumption is needed. A typical agent span, with the prompt, the response and metadata, serialises to somewhere between one and five kilobytes; long-context agents are larger. The table uses 2 KB per span, so 1 million spans is about 2 GB. Scale the data line to your own payloads.
On Pro, with 5 GB and 50,000 scores included, scoring 10% of spans with one scorer:
| Monthly spans | Data (2 KB each) | Data overage | Scores | Score overage | Total |
|---|---|---|---|---|---|
| 1,000,000 | 2 GB | none | 100,000 | 50,000 at $1.50 per 1,000 = $75 | $249 + $75 = $324 |
| 10,000,000 | 20 GB | 15 GB at $3 = $45 | 1,000,000 | 950,000 at $1.50 per 1,000 = $1,425 | $249 + $1,470 = $1,719 |
| 50,000,000 | 100 GB | 95 GB at $3 = $285 | 5,000,000 | 4,950,000 at $1.50 per 1,000 = $7,425 | $249 + $7,710 = $7,959 |
Two things jump out. The data meter is cheap: at $3 per GB, even 100 GB a month is under $300. And scoring is the bill. At 10% sampling and one scorer, scores are 80 to 90% of the Pro total; score everything with three scorers and the 10M-span month is over $40,000. The pricing rewards scoring a sample in development and penalises scoring production traffic in full.
Retention beyond 30 days is $0.50 per GB per month. Keeping the 20 GB month for a year is roughly $110 over the year, which is modest.
Where the bill surprises people
- Scores on production traffic. Turning a development scorer loose on every production span is the single biggest lever. Sample.
- Large payloads. Long-context and document-heavy agents push the data meter faster than call counts suggest.
- Model credits. Playground and judge calls draw on included credits and then on token rates. The credits are small.
- Retention. Fourteen days on Starter is short for production debugging. Thirty on Pro is workable; longer is a small add-on.
What this buys you
Braintrust is excellent at the thing it is designed for: building evaluations, running experiments across prompts and models, and keeping datasets alongside them. If your team's main loop is "change the prompt, run the eval suite, compare", it is one of the best tools available, and the unlimited-user model is fair.
Where it fits less well is production monitoring without an evaluation team: scoring every run costs money and needs scorers written, there is no schedule or heartbeat monitoring, and cost per customer is not what the meters are built around. Agent Ping's position is the opposite one: health checks on 100% of runs and AI evaluations with no scorers to write, missed-run alerts for scheduled agents, cost per customer and feature on every run, all on a flat monthly tier. The Agent Ping vs Braintrust comparison lays it out, and the LLM token cost calculator estimates the model spend inside each run.