What is traded
The orderable instrument is a model-specific bundle of token legs. The symbol represents the complete product configuration, not a generic claim on compute.
Full product identity
An instrument combines a model, a provider endpoint, a service tier, a bundle composition, and a term:
| Dimension | Meaning |
|---|---|
| Model | The model whose inference capacity is priced |
| Provider | The endpoint selected for that model and bundle |
| Service tier | The endpoint-ranking policy |
| Bundle | Exact weights across token legs |
| Term | Spot, or forward tenor plus delivery window |
Two configurations that differ on any one of these dimensions are not interchangeable. GET /api/v1/instruments returns the available symbols. POST /api/v1/bundles/quote prices a proposed bundle before any product request is submitted.
Four token legs
The public labels and API literals are:
| Product label | API literal |
|---|---|
| Prompt | PROMPT |
| Cache read | CACHE_READ |
| Cache write | CACHE_WRITE |
| Output | COMPLETION |
COMPLETION is retained as the wire literal; product copy calls the leg Output. A model or provider can leave a leg unquoted. When a requested bundle contains an unquoted leg, a quote reports the unavailable leg instead of substituting zero or another rate.
Bundle pricing
A composition assigns a non-negative decimal weight to each included leg. Valid weights sum exactly to 1, and at least two legs must carry non-zero weight.
The bundle rate is the weighted sum of the eligible endpoint's leg rates:
Weights describe how one MT or MFT of the product is apportioned; calculator workload counts are separate request inputs. Formulas includes a worked bundle example.
Endpoint selection
The service tier ranks endpoints that quote every required leg:
| Tier | Ranking |
|---|---|
CHEAPEST | Lowest bundle price |
LOWEST_LATENCY | Lowest available latency observation |
HIGHEST_THROUGHPUT | Highest available throughput observation |
Telemetry-ranked quotes may be unavailable when eligible endpoints do not publish the relevant observation. The quote response reports that state; it does not silently relabel a price-ranked endpoint as a telemetry-ranked one.