Skip to main content

Buyer guide

Braintrust Usage, Spend & Retention Checklist

Braintrust has no hard spending limit, so control the bill with spend alerts and scoring sampling rates. Size each charge from real traffic.

Reading time
3 min read
Sources
11 cited
In this piece
Braintrust+50 Strong positive

Set spend alerts and an online scoring sampling rate before production traffic reaches Braintrust, because nothing caps the bill. Charges come from the platform fee, processed data, scores, retention storage, and model credits. Size each one from a week of real traces, and check what the March 2026 plan change removed before you downgrade.

When you need this

  • You are moving from Starter to Pro, or from a pilot to production traffic.
  • Online scoring or Topics is about to run on production projects.
  • Finance asks what evals and tracing cost per month, or a project ran out of credits in the middle of an evaluation.

Evaluation criteria

  1. Estimate processed data

    How many gigabytes does a month of traffic ingest?

    Look forMeasure a week of real traces before choosing a plan. Braintrust counts every byte at ingestion across logs, experiments, and datasets, including inputs, outputs, metadata, and attachments. Starter includes 1 GB, then charges US$4 per GB. Pro includes 5 GB, then charges US$3 per GB. Deleting data or shortening retention leaves processed-data charges unchanged.

  2. Count scores

    How many scores will evals and production scoring write?

    Look forEvery offline and online score counts, including human review. Starter includes 10,000 scores a month, then charges US$2.50 per 1,000. Pro includes 50,000, then charges US$1.50 per 1,000. By GILD's arithmetic, five scorers on 10,000 production traces use up Pro's whole monthly allowance. Set a sampling rate on each online rule.

  3. Choose retention

    How long must logs stay queryable?

    Look forStarter keeps logs and playground data for 14 days. Pro keeps them for 30 days and extends to 180 days, in 30-day steps, at US$0.50 per GB a month. Braintrust bills extended retention on the month's highest retained volume. Every plan keeps experiments up to 365 days. Enterprise keeps logs up to 365 days and can delete them earlier on a schedule.

  4. Budget model credits

    What do Topics, Loop, and built-in models consume?

    Look forStarter includes US$10 in model credits a month and Pro includes US$100. Unused credits expire at month end, and token rates apply once they run out. On Starter, Topics and built-in models pause at that point unless you turn on on-demand usage. Past the credits, Topics costs US$0.06 per million input tokens and US$0.40 per million output tokens.

  5. Set alerts in place of a cap

    What stops spend from running away?

    Look forBraintrust has no hard spending limit. Set up to three spend alerts by email or Slack under Billing. The usage report splits processed data and scores from Topics tokens and retention storage. Starter also emails at 80%, 90%, and 100% of included usage.

  6. Know what the plan change moved

    Which plan are you on?

    Look forCheck legacy entitlements before downgrading. Starter replaced Free on March 16, 2026. Organizations created before then keep custom permission groups and export automations, including custom charts, until they downgrade. A downgrade removes legacy status for good.

  7. Plan for the free-tier ceiling

    What happens when a small project outgrows Starter?

    Look forDecide what happens when credits run out before you run evals: upgrade to Pro, turn on on-demand usage, or keep the work local. A developer replaced Braintrust with local storage after a June 2026 photo-judge eval exhausted write credits. A fuzzing project removed Braintrust reporting so its evaluations run without an account.

Questions to ask vendors

  1. How many gigabytes and scores did a week of production traffic produce?
  2. Which online scoring rules run, at what sampling rate?
  3. Who receives spend alerts, and at which thresholds?
  4. How long must logs be kept, and is extended retention worth its per-gigabyte cost?
  5. Are we on a legacy plan, and what would a downgrade remove?

Red flags

  • Verbose agent traces that log the full conversation on every turn.
  • Online scoring on all production traffic with several LLM-as-a-judge scorers and no sampling.
  • No spend alert on an organization with on-demand usage enabled.
  • Expecting deletion to lower this month's processed data bill.

Products in this piece

Sources

Official documentation
8
Public source
3
  1. Braintrust pricingOfficial documentationbraintrust.dev
  2. Braintrust docs: Plans and limitsOfficial documentationbraintrust.dev
  3. Braintrust docs: Billing FAQOfficial documentationbraintrust.dev
  4. Braintrust docs: Monitor usage and spend alertsOfficial documentationbraintrust.dev
  5. Braintrust docs: Configure data retentionOfficial documentationbraintrust.dev
  6. Braintrust docs: Score production tracesOfficial documentationbraintrust.dev
  7. Braintrust docs: TopicsOfficial documentationbraintrust.dev
  8. Braintrust blog: Keep building with the Starter plan (Mar 16, 2026)Official documentationbraintrust.dev
  9. julianken, bird-sight-system issue #1094: replace Braintrust after exhausting write credits (Jun 12, 2026)Public sourcegithub.com
  10. Ultrafuzz issue #1133: remove Braintrust reporting so evaluations run without an account (Sep 22, 2026)Public sourcegithub.com
  11. braintrust-sdk-javascript issue #2353: per-turn conversation logging exhausts Cloudflare Workers memory (Aug 11, 2026)Public sourcegithub.com

Keep reading

  1. Buyer guideBraintrust Evaluation & Tracing PlaybookHow to instrument an LLM feature in Braintrust and gate each release on evals, with the practices Notion and Zapier have published.
  2. Buyer guideBraintrust Data Plane, Hosting & Security ChecklistChoose a US or EU region, BYOC, or self-hosting before launch. Then move model API keys to short-lived credentials and set access and redaction.
  3. ResearchBraintrust in Practice: What Teams ReportPublic accounts from teams such as Notion, Zapier, Trainline, Dropbox and Mura report what Braintrust's evals caught, where its SDKs broke, and why teams kept it or left.
Read the research