Braintrust Pricing Explained (2026) - The Processed-Data Trap

Braintrust is the most turnkey eval platform of the major players - if blocking bad merges is your priority, nothing else is this complete out of the box. But its pricing is where teams get surprised, because the meter is unusual and it has no cap. Let me decode it, then walk a bill.

The pricing model, decoded

Braintrust has three tiers, and no per-seat charge - users are unlimited everywhere.

Tier	Price	Included	Overage	Retention
Starter	$0	$10 credits, 1 GB data, 10k scores	$4/GB data, $2.50/1k scores	14 days
Pro	$249/mo	$249 credits, 5 GB data, 50k scores	$3/GB data, $1.50/1k scores	30 days
Enterprise	Custom	Custom	Custom	Custom

The unlimited users are genuinely nice. But the quota unit is where you need to pay attention. The main meter is “processed data” in GB, and it counts every byte - inputs, outputs, prompts, metadata, traces, spans and attachments, all of it. There is a second meter for scores.

Two facts define the whole model. First, the meter punishes exactly the workloads that need observability most. A verbose multi-step agent or a RAG pipeline stuffing large contexts into every call generates a lot of bytes, and those are precisely the apps you most want to trace. Second, there is no hard spending cap. The $0 Starter jumps straight to $249/mo Pro with nothing in between, and the $249 is a floor, not a ceiling. Retention can be extended on Pro at $0.50/GB/mo.

Estimate your bill

The hard part is that processed data is measured in bytes, so your bill depends entirely on how chatty your traces are. Two examples show the range.

• A lean chatbot. Short prompts, short answers, few metadata fields. Each request might process a few kilobytes. Millions of those still fit inside a handful of GB, and you may stay near the Pro floor of $249/mo, or even inside Starter if volume is low.
• A verbose RAG agent. Every call stuffs large retrieved contexts, long system prompts and multi-step tool outputs into the trace. A single request can process hundreds of kilobytes. Push enough of those and you clear the 5 GB Pro allowance fast, then pay $3/GB on top with no cap - and the bill can climb into four figures before anyone notices.

The lever you control is trace verbosity, not request count. The same 100k requests can cost wildly different amounts depending on how much context each one carries. As one review put it, no hard spending cap means “you’ll exceed that amount without realizing it.” Model your average bytes-per-trace, not just your traffic.

The trap - no cap, and self-host is not the free escape

The uncapped meter is the headline risk. The usual escape from a usage meter is self-hosting, but with Braintrust that door is only half open.

The SDKs are open source, but Brainstore - the storage and query backend - is closed. And self-hosting is a hybrid arrangement available only on Enterprise. You run the data plane inside your own VPC via Terraform while Braintrust hosts the control plane. Your data never traverses Braintrust’s servers, which is the appeal for compliance-sensitive teams. But it is not a free, open, run-it-yourself deployment like Langfuse - you never fully own the stack, and you need an Enterprise contract to get even the hybrid version. Full detail is on the Braintrust page.

So the practical advice is simple - set up billing alerts on day one. The processed-data meter has no cap, and the workloads that need Braintrust most are the ones that run it up. The company is well-funded, with an $80M Series B in February 2026 led by Iconiq at a roughly $800M valuation, so it is not going anywhere. The risk is your bill, not their survival.

Cheaper picks, and when they win

Braintrust earns its price if turnkey CI regression gates are the job. If predictable or free pricing matters more, two alternatives win.

Langfuse is the cheaper, more predictable, open-source pick. It bills billable units at $8 per 100k - easy to model - and runs about $101/mo at 1M events managed, or free self-hosted under MIT with only three features gated. The trade-off is honest - for full regression testing you assemble the orchestration yourself, where Braintrust hands you the CI gate. If you want predictable cost and open self-host and can build a bit more of the eval pipeline, Langfuse wins.

Opik is the cheapest managed cloud. Comet’s platform is $19/mo for 100k spans, with $5 per additional 100k, and its OSS build is Apache-2.0 with the full feature set self-hosted, unlimited, no gates. It also does online evaluation, guardrails and an Agent Optimizer. If the appeal of Braintrust is the eval depth but the uncapped GB meter scares you, Opik gives you a predictable span-based bill instead. Watch the per-seat cost as your team grows.

Bottom line

Braintrust pricing is unusual in two ways that both cut the same direction - it meters processed data by the byte, and it has no spending cap. That combination punishes verbose agents and big RAG contexts, which are the apps most likely to be on it. If turnkey regression gates are worth it, use it, but turn on billing alerts before you send a single production trace. If you want predictable or free-to-self-host pricing, Langfuse or Opik at $19/mo will serve you for less.

All figures here were read from each vendor’s own materials. This category ships breaking changes monthly, so we re-verify every 30 days.