Skip to main content
Braintrust+50 Strong positive

GILD Review

GILD's call and the buying facts behind it.

GILD analysis

Braintrust Review

Pick Braintrust when evals gate releases, as at Dropbox and Fintool: a failing production trace becomes a dataset example that code, an LLM judge, or a person scores every later change against. In Parlance Labs' July 2026 test on 100 production traces, its Loop agent found the most failures of three platforms. Watch for usage with no hard spend cap, BYOC and self-hosting only on Enterprise, and the May 2026 exposure of provider keys stored on the platform.

  • AI Observability
  • AI Costs

The decision

Best fit
Teams shipping LLM features or agents that gate prompt, model, and code changes on evals and want product managers and specialists in the same tool.
Consider another workflow
For self-hosting without Enterprise, evaluate Langfuse. If your team already runs Datadog and needs production LLM monitoring, evaluate Datadog Agent Observability.
Adoption condition
Mask sensitive fields and use workload identity or least-privilege provider keys. Choose the data plane region and set spend alerts before tracing production.

Basis: Braintrust documentation, plans, and pricing, Braintrust's own incident report, and 59 public accounts from engineers and teams. These comprise 30 GitHub issues and pull requests, 6 Reddit posts and comments, 5 Hacker News comments and posts, 7 engineering and personal blog posts, 6 conference talks (2 co-presented with Braintrust), and 1 podcast episode with Braintrust's CEO. The rest are 2 independent tool evaluations, 1 X post, and 1 Product Hunt review. Parlance Labs' evaluation is listed and not counted, and the X post and the episode carry no stance. In a pilot, time a full eval run and count ingestion errors on your largest traces.

Buying & deployment factsWhat a team building on it should verify before production traffic grows.
How you buy it

US$249

A month for Pro with 5 GB and 50,000 scores. Starter is free with 1 GB.

Source 1, braintrust.dev, Pricing: PricingSource 2, Braintrust, Plans and limits: Plans and limits

Details

Starter costs US$0 monthly for 1 GB of processed data and 10,000 scores, with 14-day retention and US$10 in model credits. Pro costs US$249 for 5 GB and 50,000 scores, with 30-day retention and US$100 in credits. Every plan has unlimited users. Enterprise pricing is custom, invoiced annually.

Cost drivers

US$3 per GB

Past the Pro allowance, counted at ingestion, so deleting data does not lower it.

Source 1, braintrust.dev, Pricing: PricingSource 2, Braintrust, Plans and limits: Plans and limitsSource 5, Braintrust, Monitor usage and spend alerts: Monitor usage and spend alerts

Details

Above the allowance, processed data costs US$4 per GB on Starter and US$3 on Pro. Scores cost US$2.50 and US$1.50 per 1,000, respectively. Braintrust counts data at ingestion, so deletion does not reduce that charge. Extending Pro retention up to 180 days costs US$0.50 per GB per month. Spend alerts do not stop usage.

Where it runs

US or EU, fixed

Pick the data plane when you create the organization. Enterprise adds BYOC or self-hosting.

Source 6, braintrust.dev, Organizations and data plane region: Organizations and data plane regionSource 3, Braintrust, Deployment options: Deployment optionsSource 7, braintrust.dev, Self-hosting Braintrust: Self-hosting Braintrust

Details

Choose a US or EU data plane when creating the organization. That choice is fixed. Enterprise adds vendor-operated BYOC in your cloud account or a self-hosted data plane on AWS, GCP, or Azure. Braintrust always hosts the control plane, including login and the UI, plus project metadata.

What to verify

May 2026

An attacker reached stored provider keys in a backup. Federated identity is now supported.

Source 8, braintrust.dev, Master Subscription Agreement: Master Subscription AgreementSource 4, trust.braintrust.dev, Security incident reports: Security incident reportsSource 9, braintrust.dev, Product updates (changelog): Product updates (changelog)Source 2, Braintrust, Plans and limits: Plans and limits

Details

Braintrust's MSA prohibits AI training on customer data. In May 2026, an attacker accessed decrypted provider keys in a migration backup. Hosted organizations now support workload identity federation for Anthropic and other providers, plus assume-role for Bedrock. SAML SSO and audit logs, along with a BAA and SOC 2 report, require Enterprise.

Sourced from Braintrust documentation