Why every model sits under its cheapest listing
Nyx has one pricing rule: for every model in the catalog, our per-token price is lower than the cheapest public listing for that model on any serverless provider we track. The rule is mechanical. There is no sales motion behind it, no negotiated tier that beats it, and the current numbers are always published at /v1/models.
Tracking is a scrape, run daily. At 06:00 UTC a job pulls the public pricing pages of the eleven serverless providers we compare against, parses each into rows of model, input price, and output price, and normalizes units to dollars per million tokens. Ambiguous listings are excluded instead of guessed at: if a page quotes a bundle, a batch-only rate, or a price gated behind a quote form, it does not count. The minimum surviving price per model becomes that day's ceiling for ours.
When a listing drops, we reprice within 24 hours. The change lands in
/v1/models first and everywhere else after, and requests
bill at the price in effect when they arrive, so there is no cutover
window to reason about. This fired four times in June. The largest
move took gpt-oss-120b input from $0.031 to $0.0273 per million
overnight, because a competitor cut theirs, and our diff shipped
before their announcement post did.
Where the margin comes from
An always-lowest rule invites an obvious suspicion, that the prices are investor-funded and will snap back later. The honest answer is in the unit economics. Nyx serves a catalog of ten open models at high utilization instead of a long tail at low utilization: batches stay dense, prefix caches stay hot, and measured throughput per GPU on the catalog models runs about 2.1x a reference vLLM deployment. The 71µs serving path is part of the same posture, since time a request does not spend in routing is capacity the fleet can sell. When cost per token sits well below the market's cheapest listing, pricing under that listing still leaves margin. It is a thinner margin than the market leader takes, and it is still a margin.
The rule also has an honest edge case. If a price war pushed a listing below our serving cost for some model, we would delist that model and say so rather than serve it at a loss to keep a slogan intact. It has not happened. Over the six months the scrape has run, the gap between our cost and the cheapest listing has widened, because serving efficiency compounds and list prices move in steps.
The practical summary for a buyer is short. Read /v1/models, compare it against any public listing you can find, and hold us to the rule. The scrape runs whether anyone is watching or not.
← All posts