Skip to main content
The Workload Cost Index is the OTPI companion for the price of a call. OTPI prices the token; this prices the request:
Where OTPI blends posted per-token prices, the Workload Cost Index tracks the settled dollars a lab’s paid workload actually generated per request that day, plus the two things that drive it: same-model unit cost (mix-purged) and tokens per request. It is published for the three closed labs — Anthropic, OpenAI, and Google — and refreshed on the same daily settle as OTPI.

​
The three series

Every workload node (a provider or a model class within it) publishes three daily series: Read the three together: if Cost per Request falls while the Like-for-Like Index stays flat, the lab is getting cheaper by shifting workload to smaller models rather than because the same model got cheaper. If the Like-for-Like Index also falls, the same model actually got cheaper. Rising Workload Intensity — longer answers, more tool calls, deeper reasoning — can push Cost per Request up even while per-token prices fall.
“Settled” here means the index’s official daily computation from venue-reported realized prices and volumes — the same epistemic status as OTPI. It is not audited payments.

​
Providers and model classes

Model classes are exact partitions of a provider’s daily per-model rows, so class-level rows sum back to the provider aggregate. Class-level like-for-like removes only within-class mix. Choose the aggregate for a whole-lab view, or a class to isolate a family (for example sonnet to strip out Opus/Haiku mix effects).
Thin classes (OpenAI pro, codex, nano, and the image families across labs) are accurate partitions but low-volume. Daily wiggles in those series are small-sample noise and should be read with wider windows.

​
How it’s computed

The index rides the existing OTPI settle. For each provider on each settlement day:
  1. Per-model daily rows are persisted with settled dollars, tokens, and request counts.
  2. Provider and class nodes are built by summing the partition of models in each node.
  3. Cost per Request is settled_dollars ÷ requests on that day.
  4. Workload Intensity is tokens ÷ requests on that day.
  5. Like-for-Like Index uses an LMDI decomposition between trailing 7-day windows to separate the same-model unit-cost change from the mix change, then chains the same-model component day over day. Inputs are smoothed, outputs are not, so daily chaining does not drift and model entry/exit lands in the mix term by construction.
Errors in the workload settle are isolated from the OTPI settle — OTPI rows are written first, then workload errors are collected and thrown as an AggregateError so the job exits with failure while preserving the token price index.

​
Coverage window

  • History starts 2026-06-11. The 2026-06-04 to 2026-06-10 settlement block was backfilled with elevated Opus effective prices and is excluded from the seeded base.
  • 2026-06-30 is missing from settlement inputs for all labs — a one-day gap pending a settlement re-emit upstream.
  • Newly settled days advance daily on the same ~36-hour lag as OTPI.

​
Accessing the data

Query the index programmatically via GET /api/workload. Unlike OTPI, this endpoint has no free tier — every request must include an API key. See Authentication for how to send the Authorization: Bearer header, and Manage API keys to create one.

​
Viewing in the dashboard

The Workload Cost tab on data.ornn.com/analytics plots all three series, with dropdowns for provider and model class. Axes and hover tooltips are unit-aware — cents for cost per request, index points for like-for-like, thousands of tokens for intensity.