Braintrust
The deepest eval-to-CI loop in the category, and the only platform here that bills every byte of your traces with no spending cap, so set billing alerts before the first production trace.
PRICING
| Starter | $0 | $10 model credits, 1 GB processed data, 10,000 scores, 14-day retention, unlimited users, projects, datasets, playgrounds and experiments. One human review score config per project. Overage $4/GB and $2.50 per 1,000 scores. No credit card required |
| Pro | $249/mo | $249 model credits, 5 GB processed data, 50,000 scores, 30-day retention then $0.50/GB/month, unlimited human review scores, custom charts, environments, S3 export, basic RBAC, priority support. Overage $3/GB and $1.50 per 1,000 scores. 6-12 months free for qualifying startups |
| Enterprise | Custom | Custom retention and export, custom RBAC, SAML SSO, DPA and HIPAA BAA, uptime SLA, shared Slack channel, and the only route to hybrid self-hosted or BYOC deployment |
The meter is the thing to understand before signing. Braintrust bills processed data in GB, counting every byte of inputs, outputs, prompts, metadata, spans and attachments, plus a second meter on scores. There is no hard spending cap and nothing between $0 Starter and $249 Pro. Verbose multi-step agents and large RAG contexts, the workloads that most need tracing, burn the allowance fastest, and a team running five scorers per output hits the score cap five times faster than a team running one. Model your average bytes per trace, not your request count. Euro figures converted from published USD at roughly 0.92.
checked 2026-08-10 · vendor pricing page
Element scores
Strengths
Braintrust is the eval platform other tools bolt evaluation onto afterwards. The loop is genuinely closed: a production trace becomes a dataset row in one click, the dataset feeds an immutable experiment with side-by-side diffs, a GitHub Action runs that experiment on every pull request and blocks the merge when quality drops, and the same scorers then run online against live traffic. Autoevals ships working scorers out of the box, code and LLM-as-judge and human review sit in one system, and Loop turns the tedious half of eval work into natural language: analyse the logs, generate the SQL filter, build the dataset, write the scorer, optimise the prompt. Tracing underneath is properly typed, with nested spans for tool calls, reasoning steps, state and memory, wrappers for more than 20 agent frameworks, an OpenTelemetry processor, and Brainstore as a purpose-built query engine. Multimodal is handled well, with base64 payloads from OpenAI, Anthropic, Google and Bedrock auto-extracted into previewable attachments. The MCP server runs on both a US and an EU data plane. And the commercial shape suits teams: no per-seat charge anywhere, unlimited users even on the free tier, and $80M raised in February 2026 at an $800M valuation, so it is not going anywhere.
Honest dings
The billing meter is the risk, and it is a real one. Braintrust charges by processed data in GB, counting every byte of every trace, with a second meter on scores and no hard spending cap. The jump is $0 to $249 per month with nothing between, and the workloads with the most to gain from tracing, verbose multi-step agents and large RAG contexts, are exactly the ones that burn GB fastest. Turn on billing alerts on day one. Second, self-hosting is half a door: the SDKs are open but Brainstore is closed, and the only deployment you can own is an Enterprise hybrid where you run the data plane via Terraform while Braintrust keeps the control plane. There is no free self-host, and SOC 2, HIPAA, SSO and custom RBAC all sit behind that same Enterprise conversation, so a small regulated team gets priced into a sales cycle. Third, the eval surface has honest gaps: experiment results carry no statistical significance, so you see a score move without knowing whether it means anything; there are no native multi-turn conversation evals, and teams work around it by concatenating traces from metadata; and there is no rules-based routing from production traces into datasets or annotation queues, so triage stays manual. The free tier's 14-day retention is too short to compare experiments across sprints, and it allows only one human review score config per project.
Sources (13) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- braintrust.dev/pricing — Official pricing, verified live 10/08/2026: Starter $0 with $10 credits, 1 GB processed data, 10k scores, 14-day retention; Pro $249/mo with $249 credits, 5 GB, 50k scores, 30-day retention then $0.50/GB/mo; overage rates; unlimited users on every tier; SSO, custom RBAC, BAA and uptime SLA all Enterprise-only (accessed 2026-08-10)
- braintrust.dev/docs/self-hosting — Official self-hosting docs and the source of the hybrid finding: the customer runs the data plane (API, Postgres, Redis, object storage, Brainstore) via Terraform on AWS, GCP or Azure while Braintrust retains the control plane, auth and metadata. Includes the table of exactly which data sits where (accessed 2026-08-10)
- braintrust.dev/docs/reference/mcp — Official MCP docs: native server at api.braintrust.dev/mcp with an EU data plane at api-eu.braintrust.dev/mcp, API-key or OAuth auth, setup for Claude Code, Claude Desktop, Codex, Cursor, VS Code and Devin. Also states the bt CLI is the recommended interface for coding agents (accessed 2026-08-10)
- braintrust.dev/docs/guides/loop — Official Loop documentation: model selection across Claude, GPT and Gemini limited to org-level providers, auto-accept toggle, @-mention data sources, slash commands including /generate-filter and /find-similar, dataset and scorer generation from natural language. Primary evidence for the Agents and Synthetic data scores (accessed 2026-08-10)
- braintrust.dev/docs/instrument/attachments — Official multimodal evidence: attachments for images, audio, video, PDF and large JSON, automatic extraction of inline base64 from OpenAI, Anthropic, Google inline_data and Bedrock Converse content blocks, inline preview, and the documented limitation that PDFs cannot render inside custom views (accessed 2026-08-10)
- braintrust.dev/docs/evaluate — Official eval lifecycle: playgrounds to experiments to CI to online scoring to dataset feedback, offline versus online evaluation, code scorers, LLM-as-a-judge and classifiers (accessed 2026-08-10)
- braintrust.dev/articles/agent-observability-complet… — Official, 21/06/2026: free tier limits, named production customers (Notion, Stripe, Vercel, Zapier, Airtable, Instacart), the Notion 3-to-30 issues per day figure, framework wrapper list, OpenTelemetry span processor and the vendor's own competitive table against Galileo, Arize Phoenix and Datadog (accessed 2026-08-10)
- axios.com/pro/enterprise-software-deals/2026/02/17/… — Independent confirmation of the $80M Series B led by Iconiq at an $800M post-money valuation, sourced directly to CEO Ankur Goyal, 17/02/2026 (accessed 2026-08-10)
- siliconangle.com/2026/02/17/braintrust-lands-80m-se… — Second independent report of the Series B: Braintrust Data Inc., San Francisco, investors Iconiq, Andreessen Horowitz, Greylock, basecase capital and Elad Gil, funding earmarked for engineering, GTM and new regions (accessed 2026-08-10)
- llmtools.cc/blog/braintrust-pricing — Independent pricing analysis (26/07/2026) and the source of the uncapped-meter finding: processed data counts every byte of inputs, outputs, prompts, metadata and attachments, no hard spending cap, $0 jumps straight to $249, and verbose agents or large RAG contexts burn the allowance fastest. Also documents that Brainstore is closed and self-host is Enterprise-only hybrid (accessed 2026-08-10)
- aitoolsbakery.com/blog/braintrust-review — Independent hands-on review (09/06/2026): confirms the free tier is genuinely usable, that cost scales with scores rather than traces so five scorers per output reaches the cap five times faster, and that self-hosting plus strong governance controls are Enterprise-tier only (accessed 2026-08-10)
- coverge.ai/blog/braintrust-pricing — Independent tier-by-tier breakdown (14/04/2026) and the source of the disputed-gating point: states Braintrust scores a pipeline but has no mechanism to automatically block a bad version from reaching production, no human approval gate between eval and deploy, and no structured compliance proof bundles. Also confirms 1 GB is roughly one million trace spans (accessed 2026-08-10)
- langchain.com/resources/langsmith-vs-braintrust — Competitor-authored comparison (LangChain, 06/06/2026), treated as adversarial but specific and checkable: no native trace-to-dataset or trace-to-queue automation, no multi-turn context or thread-level evals, no statistical significance on experiment results, review workflow is score-only, and one review scorer per project on Starter (accessed 2026-08-10)