Costs & Pricing

Every LLM call traced by Traccia carries cost information. The platform maintains a two-layer cost model: an SDK-reported estimate emitted at run time, and a platform-recomputed authoritative cost calculated server-side using the platform's own up-to-date pricing snapshot and any org-level overrides you have configured. Platform costs are what you see throughout the Traccia dashboard.

The Two-Layer Cost Model

Each LLM span carries two separate cost values. They are always kept distinct — the SDK value is never overwritten.

Where Costs Appear

Platform-recomputed costs are the primary figure shown everywhere in the dashboard. SDK-estimated costs are available as a secondary detail for comparison and debugging.

When a span does not yet have a platform-recomputed cost (for example, immediately after ingestion before the background annotation job has run), the SDK estimate is shown as the fallback.

Platform Pricing Sync

The platform maintains its own pricing snapshot, sourced from a continuously maintained upstream database of 2,500+ models. It is refreshed automatically every 6 hours. The sync sequence is:

In addition to the sync-triggered backfill, a live annotation job runs every 15 seconds and annotates any recently-ingested spans that do not yet have a platform cost. This means new traces are typically annotated within 15 seconds of ingestion.

Org-Level Pricing Overrides

Org admins can define custom per-model prices that override the platform's global snapshot for all agents in the organisation. Common use cases include fine-tuned models with custom API pricing, negotiated enterprise rates, or internal cost allocation that differs from list prices.

Setting an override

1. 1

   Go to Settings → Pricing in the Traccia dashboard.

2. 2

   Find the model you want to override, or add a new entry for a custom model name.

3. 3

   Enter prompt and completion rates in USD per 1,000 tokens. Optionally set cache_write and cached_prompt rates for models that support prompt caching.

4. 4

   Save. You will be offered the option to "Also recompute past traces" — enable this to retroactively apply the new rates to existing spans.

Override JSON schema

Whether set via the UI or programmatically, all pricing overrides use the same JSON shape:

{
  "<model-name>": {
    "prompt": 0.005,
    "completion": 0.015,
    "cache_write": 0.001,
    "cached_prompt": 0.0005
  }
}

All rates are USD per 1,000 tokens. cache_write and cached_prompt are optional.

Recomputing Past Traces

When you save a pricing override, the save dialog offers a "Also recompute past traces" option. When enabled:

• All LLM spans for models covered by the override are re-evaluated against the new rates.
• platform_cost_usd and spans.cost_usd are updated in bulk.
• Trace-level totals (traces.total_cost) are recalculated to reflect the updated span costs.
• Each updated span records which pricing version was used for the recomputation, visible in the span detail panel.
• The original llm.cost.usd SDK value is never overwritten.

Recomputation scope defaults to the past 24 hours. For longer lookback windows, contact support.

When SDK and Platform Costs Diverge

You may occasionally see a muted "SDK reported $X.XXXX" note alongside the platform cost in the span detail panel. This is expected and happens for one of three reasons:

Cost Attribution Dimensions

The platform attributes costs along several dimensions for filtering and analysis: