Spreads
What several providers charge for the same product, cheapest first, with the percentage between the low and the high.
GET /api/v1/spreads/{class}/{product}curl https://vitna.ai/api/v1/spreads/compute/h100-sxm
curl "https://vitna.ai/api/v1/spreads/model/kimi-k3?dimension=output_per_mtok"Parameters
| Parameter | In | Meaning |
|---|---|---|
class | path | model, tooling, retrieval or compute |
product | path | Product slug, lowercased and dashed |
dimension | query | Which meter. Optional. |
Product slugs come from the sku's attrs, not from a separate table. H100 SXM is h100-sxm.
Default dimension
Omitting dimension gives the one the boards draw:
| Class | Default |
|---|---|
model | input_per_mtok |
tooling | tool_call_per_1k |
retrieval | embed_per_mtok |
extraction | page_ocr |
compute | gpu_hour_on_demand |
Response
{
"class": "compute",
"product": "H100 SXM",
"productSlug": "h100-sxm",
"dimension": "gpu_hour_on_demand",
"unit": "USD per GPU-hour",
"providerCount": 6,
"minUsd": 1.99,
"maxUsd": 3.29,
"spreadPct": 65,
"freshestAt": "2026-08-23T08:00:11.402Z",
"stalestAt": "2026-08-22T19:00:04.118Z",
"anyStale": false,
"providers": [ { "provider": "...", "skuSlug": "...", "amountUsd": 1.99, "derived": false, "stale": false, "fetchedAt": "...", "evidenceSha256": "..." } ],
"dimensions": ["gpu_hour_on_demand"],
"meta": { "generatedAt": "...", "source": "database" }
}spreadPct is (max - min) / min, as a rounded whole percent. It is expressed against the cheapest seller, so 65 means the dearest costs 65 percent more than the cheapest, not that it is 65 percent of it.
dimensions lists every meter this product has a spread for, so a caller can discover what else to ask for without a second request.
Two providers minimum
A dimension with fewer than two priced providers has no spread and is omitted. A single-seller meter is a price, not a spread, and reporting a zero percent spread for it would suggest agreement between sellers that do not exist.
Cache writes
A spread on cache_write_5m_per_mtok or cache_write_1h_per_mtok carries a blockNote. Sellers cache in blocks of a size each one chooses, so two equal rates per million tokens at one retention window need not buy the same caching. No other meter carries the field. See what a cache write can and cannot be set against.
Errors
| Status | Code | Cause |
|---|---|---|
| 404 | unknown_class | Not a registered class |
| 404 | unknown_product | No product by that slug in the class |
| 404 | unknown_dimension | Product exists, is not metered that way. The message lists the ones it has. |
Two paths, one shape
With a database the answer comes from the SQL view; without one it is computed from the committed snapshot. The shapes are identical by design, so the site and the API cannot disagree about a figure. meta.source is the only way to tell which path answered.
Caching
public, max-age=60, stale-while-revalidate=3600.
