What it is
Braintrust combines evaluation and observability. You run experiments that compare prompts and models against datasets, score them with code, LLM judges or human review, and then log production traces and score them online with the same scorers. It also offers a prompt playground, an AI gateway for routing model calls, an MCP server, a bt command-line tool and Loop, an AI agent that analyses logs and drafts evals.
Traces arrive through SDKs and integrations for OpenAI, Anthropic, LangChain, LlamaIndex, Pydantic AI, Google ADK and others, or through OpenTelemetry. The SDKs can redact sensitive data before upload. Its open-source autoevals library (MIT) supplies common scorers.
The free Starter plan includes 1 GB of processed data, 10,000 scores, $10 of model credits and 14-day retention with unlimited users. Pro costs $249 a month for 5 GB, 50,000 scores, $100 of model credits and 30-day retention (extendable to 180 days at $0.50 per GB a month). Self-hosting the data plane, SAML SSO and HIPAA agreements are Enterprise-only. Braintrust raised an $80 million Series B in February 2026.
Our review
Braintrust puts evaluation first and tracing second, and that order suits teams whose real problem is knowing whether a prompt or model change helped. Experiments, datasets, scorers and online scoring share one data model, so a failing production trace becomes a test case in a few clicks, and the same scorer runs offline in CI and online against live traffic.
Pricing is by processed data and scores, with unlimited users on every plan. That is friendlier to a 40-person product organisation than LangSmith's per-seat model, and less friendly to an agent that writes long traces. Pro's $249 a month covers 5 GB and 50,000 scores; beyond that, each GB costs $3 and each 1,000 scores $1.50, so scoring every production request with an LLM judge is the line that grows fastest. Retention is short by default, 14 days on Starter and 30 on Pro, with 180 days available on Pro at $0.50 per GB a month.
Data control is the dividing line. The server is not open source, and keeping traces in your own cloud means an Enterprise contract. The hybrid design is sound: according to the self-hosting docs, the data plane (API, Postgres, Redis, object storage and the Brainstore engine) runs in your AWS, GCP or Azure account from official Terraform modules, SDKs send data straight to it, and Braintrust's hosted control plane holds only metadata such as project names, hashed API keys and login details. Health and billing telemetry still flow back by default. Braintrust's security page states SOC 2 Type II and offers BAAs, and itself recommends a self-hosted or BYOC deployment in the EU for full GDPR compliance.
Instrumentation carries little lock-in, because OpenTelemetry ingestion and the MIT-licensed autoevals scorers work elsewhere. The lock-in is in the eval history, datasets and online scorers a team builds up, which is precisely the part Braintrust does better than Langfuse or Arize Phoenix, and precisely the part that is hardest to move.
Where it earns its place
A six-engineer team shipping a customer-support agent
Stop prompt changes from quietly breaking answers
A golden dataset of past tickets runs as an experiment on every pull request, scored by autoevals and an LLM judge; regressions block the merge. On Pro at $249 a month, 50,000 scores covers CI plus sampled production scoring.
An ML platform lead at a fintech with EU customers
Trace LLM calls without sending customer data to a US SaaS
An Enterprise contract deploys the data plane into the company's own AWS or GCP account with Terraform; traces, prompts and datasets stay in-region under its own KMS keys, while Braintrust hosts only the UI and metadata.
A head of product at a 40-person AI startup
Give PMs and engineers one place to review model quality
Unlimited users means no per-seat negotiation; PMs add human review scores to traces while engineers run experiments. The bill tracks data volume, so trimming large payloads matters more than headcount.
Before you commit
- Setup
- Create a project, install the Python or TypeScript SDK, wrap the model client and run a first Eval() against a small dataset; existing OpenTelemetry exporters can point at Braintrust instead.
- Integrations
- SDKs and wrappers for OpenAI, Anthropic, LangChain, LlamaIndex, Pydantic AI and Google ADK; OpenTelemetry; AI gateway; MCP server; bt CLI.
- Data and privacy
- SOC 2 Type II; data encrypted at rest and in transit; no published statement on training use; configurable retention policies; self-hosted data plane, SAML and BAA on Enterprise only.
- Cost at scale
- A production agent logging 20 GB a month on Pro pays $249 plus 15 GB at $3, about $294, before extra scores at $1.50 per 1,000 and retention beyond 30 days at $0.50 per GB a month.
- Limits
- Starter keeps data 14 days and allows one human review score per project; Pro keeps 30 days unless extended; custom charts and environments need Pro.
What's good
- Evals, datasets and online scoring share one data model
- Unlimited users on every plan, including free
- Hybrid self-hosting keeps traces in your AWS, GCP or Azure account
- OpenTelemetry ingestion and open-source autoevals scorers
- SOC 2 Type II; BAAs available on Enterprise
What's not
- Pro starts at $249 a month, a big step from free
- LLM-judge scoring of production traffic adds $1.50 per 1,000 scores on Pro
- Default retention of 14 or 30 days
- Server not open source; self-hosting and SAML only on Enterprise
Best for
Product teams who want evaluation-driven development with hosted tracing and are happy to pay for a managed service.
Alternatives
Braintrust: common questions
- Is Braintrust worth it?
- Braintrust is built for teams whose hardest question is whether the last change made the product better, and it charges by data rather than by head, which flatters big teams and punishes chatty agents. We rate it 4.3 out of 5.
- How much does Braintrust cost?
- Starter free (1 GB data, 10k scores, $10 model credits, 14-day retention, unlimited users; then $4/GB and $2.50 per 1,000 scores). Pro $249/month (5 GB, 50k scores, $100 model credits, 30-day retention; then $3/GB and $1.50 per 1,000 scores). Enterprise custom, with SAML, BAA and self-hosted data plane. Qualifying startups get Pro free for 6 to 12 months.
- What are the downsides of Braintrust?
- Pro starts at $249 a month, a big step from free; lLM-judge scoring of production traffic adds $1.50 per 1,000 scores on Pro; default retention of 14 or 30 days; server not open source; self-hosting and SAML only on Enterprise.
- Who is Braintrust best for?
- Product teams who want evaluation-driven development with hosted tracing and are happy to pay for a managed service.
- How much does Braintrust cost for a team or in production?
- A production agent logging 20 GB a month on Pro pays $249 plus 15 GB at $3, about $294, before extra scores at $1.50 per 1,000 and retention beyond 30 days at $0.50 per GB a month.
- Is Braintrust safe for company data?
- SOC 2 Type II; data encrypted at rest and in transit; no published statement on training use; configurable retention policies; self-hosted data plane, SAML and BAA on Enterprise only.
- What are the best Braintrust alternatives?
- The closest alternatives we have reviewed are LangSmith.
Answers checked Oct 23, 2026.
Make Braintrust? Show this rating on your site
