Models & Pricing
Every cost number in the platform comes from a model definition — the rule that says how to recognize a model in your traces and what it charges per token. Keeping these accurate is what makes FinOps trustworthy.
Manage model definitions under Settings → Models. The platform ships with built-in definitions for common models; add custom ones for anything it doesn't know.
What a model definition contains
| Field | Purpose |
|---|---|
| Match pattern | A pattern that identifies the model name in incoming generations, so cost is applied to the right calls. |
| Pricing | Input and output token prices, used to compute cost. |
| Tokenizer | How tokens are counted (important when usage isn't reported by the provider). |
| Maintainer | Built-in (platform-provided) or custom (yours). |
How cost is computed
For each generation, cost = (input tokens × input price) + (output tokens × output price), using the matched definition's pricing. If your trace already includes usage and cost from the provider, that's used directly; otherwise the platform applies your model definition.
Built-in vs. custom models
- Built-in models are maintained for you and are read-only. You can clone one to customize its pricing.
- Custom models cover anything the platform doesn't know about — self-hosted, fine-tuned, or newly released models. Define the match pattern, pricing, and tokenizer.
Check coverage
In Settings → Models, confirm the models in your traces have definitions. Generations with no matching definition won't have accurate cost.
Add a custom model
For an unrecognized model, create a definition: match pattern, input/output prices, and tokenizer.
Verify
Send a trace using that model and confirm cost appears correctly in AI Usage.
If costs look wrong or missing, it's almost always a model-definition gap — an unmatched model name or stale pricing. Start your debugging here.
Next steps
- AI Usage — see the cost your definitions produce.
- AI Optimizer — accurate pricing makes savings estimates reliable.
- Sending Data In — include model and token details for best results.