Logfire is the production AI improvement system: investigate the complete agent and application trace, annotate the runs that matter, score production and offline datasets with Pydantic Evals, let the optimizer propose a trace-backed change, then ship managed prompts, agent specs, tools, and skills with targeting and rollout controls. Add infrastructure monitoring, browser session replay, feature flags, and the agent trace investigator—all in one OpenTelemetry-native product, without a separate per-score meter.
Already use Braintrust’s Eval API? Change two environment variables and send your next runs to Logfire’s Evals workspace. Your inline data, local tasks, and scorers can stay where they are.
export BRAINTRUST_APP_URL="https://logfire-us.pydantic.dev/v1/braintrust"export BRAINTRUST_API_KEY="<your-logfire-write-token>"Use https://logfire-eu.pydantic.dev/v1/braintrust for an EU project. The endpoint does not import existing history or provide Braintrust-hosted datasets, prompts, functions, a model proxy, server-side scoring, or public sharing.
| Feature | Logfire | Braintrust |
|---|---|---|
| Production context | Browser, agents, services, databases, logs, metrics, and infrastructure | AI and application traces, logs, and OpenTelemetry spans |
| Full-stack observability suite | Infrastructure monitoring, service maps, logs, metrics, traces, and AI | No |
| Browser session replay | Browser tracing and session replay alongside the agent trace | No |
| First-class feature flags | OpenFeature/OFREP flags, targeting, and controlled rollout | No |
| Evaluation workflow | Pydantic Evals: the same evaluators online and offline | Experiments, playgrounds, CI/CD, and online scoring |
| Score pricing | No separate score meter; normal record pricing applies | $1.50 per 1,000 scores after 50K/month on Pro |
| Human review | Annotation queues on production runs and evals | Human-review scores and assigned trace review |
| From failure to change | Trace-backed optimizer and managed agent configuration | Prompt, scorer, dataset, and environment workflows |
| Managed agent configuration | Prompts, agent specs, tools, skills, versioning, targeting, and rollout | No |
| Controlled rollout | Immutable versions, labels, targeting, weighted rollout, and feature flags | Prompt, dataset, and parameter environments |
| Investigation workflow | Agent trace investigator, PostgreSQL-compatible SQL, and MCP across full telemetry | Logs, trace views, SQL, and MCP for Braintrust data |
| Deployment options | Cloud, Dedicated, or the same product self-hosted on Kubernetes | Cloud or an Enterprise self-hosted data plane |
| Production coverage | Scores | Braintrust Pro base + score charge | Estimated Logfire monthly charge |
|---|---|---|---|
| 10M runs × 10% sampled × 3 scores | 3M scores/month | $4,674/month | $49/month |
| 10M runs × 25% sampled × 5 scores | 12.5M scores/month | $18,924/month | $54/month |
| 100M runs × 10% sampled × 5 scores | 50M scores/month | $75,174/month | $129/month |
Calculation scope
Braintrust Pro base plus score charges compared with the Logfire plan and normal record usage attributable to the score results. Processed-data and extended-retention charges are disclosed but omitted from the totals because the score count does not determine either meter.
Basis: USD; public monthly list price. Prices checked .
3,000,000 recorded scores in the month.
$249 $249 (3,000,000 − 50,000 included) ÷ 1,000 × $1.50 $4,425 $49 $49 (3M − 10M included) ÷ 1M × $2 $0 12,500,000 recorded scores in the month.
$249 $249 (12,500,000 − 50,000 included) ÷ 1,000 × $1.50 $18,675 $49 $49 (12.5M − 10M included) ÷ 1M × $2 $5 50,000,000 recorded scores in the month.
$249 $249 (50,000,000 − 50,000 included) ÷ 1,000 × $1.50 $74,925 $49 $49 (50M − 10M included) ÷ 1M × $2 $80 “50k scores + $1.50/1k”
Braintrust's on-page calculator itemizes its platform, processed-data, score, and retention meters.
“All plans include unlimited users, projects, experiments, and datasets.”
Braintrust Pro includes $249 per month in model credits, 5 GB processed data, and 30-day retention.
“Covers 10 million logs, spans, and metrics every month.”
The official calculator shows 20 million records and no extra seats totaling $69/month on Team.
An agent failure is often a browser, API, database, retrieval, tool, or infrastructure failure wearing an LLM-shaped mask. Logfire keeps those signals in one nested trace, with service maps, logs, metrics, SQL, and an agent trace investigator built for the production incident behind the score.
Use the same Pydantic Evals evaluators for fast offline feedback and online production monitoring. Cheap heuristics can run on every run; LLM judges can sample the traffic that deserves them. Scores use Logfire's normal record pricing instead of adding a new billing meter.
Annotation queues let reviewers work through the production runs that matter, with verdicts, failure categories, expected outputs, comments, and tags. That judgment stays linked to the trace, becomes a reusable evaluation case, and gives the optimizer grounded evidence for the next change.
Logfire manages prompts, agent specs, tools, and skills as versioned configuration. Review a trace-backed proposal, then target a cohort, canary a weighted rollout, watch the live result, and roll back by moving a label. The version that served every run is part of that run's trace.
Yes. Point the Braintrust SDK at Logfire and keep your existing eval suite. Future eval runs appear in the Evals workspace alongside the traces and telemetry that explain each result. Existing Braintrust history and Braintrust-hosted services are not imported.
Logfire does not add a separate per-score meter: score results use its normal record allowance and $2-per-million paid-plan overage. In the worked model, 3 million score records stay within Personal at $49; 12.5 million require Team and total $54; and 50 million total $129. Braintrust Pro lists a $249 monthly base with 50,000 scores included, then $1.50 per 1,000 scores, producing $4,674, $18,924, and $75,174 for the same three score counts. Pro also includes $249 in model credits; those credits do not reduce the invoice but can offset built-in model usage. Processed-data and extended-retention charges are excluded.
Yes. Both support offline experiments and asynchronous scoring of production traffic. Logfire uses the same Pydantic Evals evaluators online and offline, keeps every result attached to its OpenTelemetry trace, and does not add a separate score meter.
Logfire investigates the whole production system, not only the AI output: browser activity, agent and model calls, tool calls, APIs, databases, logs, metrics, services, and infrastructure. From a failing run, teams can annotate it, query it with SQL or MCP, export it into an evaluation case, use the prompt optimizer to propose a trace-backed change, and ship managed prompts, agent specs, tools, and skills with versioning, targeting, and rollout controls.
Yes. Logfire combines online and offline Pydantic Evals, annotation queues, no separate per-score meter, trace-backed prompt optimization, and managed configuration with the full production context around an agent: browser, services, databases, logs, metrics, and infrastructure.
Get started with 10 million free spans, logs, and metrics per month. No credit card required.