Configure custom model costs for estimation

Applies to:

Plan - Any
Deployment - Any

Summary

If you don’t log metrics.estimated_cost directly on a span, Braintrust falls back to the model registry to estimate cost from token metrics and registered pricing. Custom LLM models need to be registered in Braintrust’s model registry with pricing information for this fallback to produce a cost. Spans that log metrics.estimated_cost explicitly use that value as-is in the trace viewer and in the estimated_cost() SQL function.

Configuration Steps

Step 1: Register your custom model

Register custom model pricing on an organization-level provider. The pricing lookup used by estimated_cost() does not include project-level providers. Go to Settings > Models. Click Organization provider, then choose Custom providers > New. On Setup, configure the provider, then under Models, click Custom model and enter the model identifier recorded in your traces. See Custom providers.

Step 2: Set pricing information

Click the settings icon beside the custom model to open Model options, then configure:

Input cost per million tokens
Output cost per million tokens
Cache read/write costs (if using prompt caching)

Click Create, or Update if you are editing an existing provider, to save the configuration.

Step 3: Update span metadata

Ensure your logged spans include metadata.model matching the exact registered model name.

braintrust.log(
    metadata={"model": "your-custom-model-name"},
    # other span data
)

Step 4: Query costs across projects

Use SQL to aggregate costs across multiple projects:

SELECT
    metadata.model,
    day(created) AS date,
    avg(estimated_cost()) AS avg_cost,
    sum(estimated_cost()) AS total_cost
FROM project_logs('project-id-1', 'project-id-2')
WHERE created > now() - interval 7 day
  AND span_attributes.type = 'llm'
GROUP BY metadata.model, date

Save this query as a custom view for reuse across your organization. For more granular breakdowns — for example, separating cached prompt cost from uncached prompt cost — use estimated_cost_component(name) to extract a single cost component, or estimated_cost_breakdown() to return all components as a JSON object. See SQL functions for the available component names.

Log a discounted cost

If your provider applies a discount or batch pricing that Braintrust’s model prices do not reflect, calculate the discounted cost (using the provider’s applicable rates, including cached-token pricing where relevant) and log it in USD as metrics.estimated_cost on the corresponding LLM span. Supply the final cost, not the discount percentage. This overrides automatic estimation without requiring a custom model. For example, a request costing $0.003548 at standard rates costs $0.001774 with a 50% discount. Log 0.001774 as metrics.estimated_cost. Use the provider’s applicable rates, including cached-token pricing where relevant. For existing spans, see Update spans. You need the LLM span’s id and must ensure the original span is flushed before updating it.

​Summary

​Configuration Steps

​Step 1: Register your custom model

​Step 2: Set pricing information

​Step 3: Update span metadata

​Step 4: Query costs across projects

​Log a discounted cost

Summary

Configuration Steps

Step 1: Register your custom model

Step 2: Set pricing information

Step 3: Update span metadata

Step 4: Query costs across projects

Log a discounted cost