Inference · measured in public
The models this site runs on, measured like a provider table
Hugging Face publishes a table of inference providers: for each model, what a million tokens costs, how big the context window is, how fast the first token arrives, and how quickly the rest stream. This page points that instrument at ourselves. These are the models this site actually routes traffic to, measured by our own probe, with the raw trials published underneath every number.
Heads up: sample data. The timings below are from a zero-cost practice run with stand-in numbers, so the table has something to show while the probe gets its first real run. Prices and context windows are real; TTFT and throughput aren't measurements yet.
The roster
Every model here is one the site really uses; the drift guard in CI fails if this table and the /chat allowlist or the /council panel ever disagree. Click a column to sort; click a row for the raw trials.
✓ re-derived in your browser: every displayed number was just recomputed from the raw trials below with the same code CI runs.
| Status | ||||||
|---|---|---|---|---|---|---|
| $15.00 | $75.00 | 200,000 | 0.97 s | 34 tok/s | ● live | |
| $3.00 | $15.00 | 200,000 | 0.67 s | 60 tok/s | ● live | |
| $1.00 | $5.00 | 200,000 | 0.41 s | 110 tok/s | ● live | |
| n/a | n/a | 131,072 | 0.75 s | 36 tok/s | ● live | |
| n/a | n/a | 131,072 | 0.75 s | 38 tok/s | ● live | |
| n/a | n/a | 131,072 | 0.85 s | 29 tok/s | ● live | |
| n/a | n/a | 131,072 | 2.35 s | 28 tok/s | ● live | |
| n/a | n/a | 32,768 | 0.48 s | 85 tok/s | ● live |
Prices: committed vendor list prices ($/1M tokens). “n/a” on open models is honest: the HF Router bills pass-through at whichever provider it picks per request, so there is no single true number to print. TTFT and throughput are medians across the probe's trials; click a row for the raw receipts.
8 models · 3 trials each · probed 2026-07-27 · prices as of 2026-06 · catalog as of 2026-07.
Methodology
How this is measured
One prompt, every model
The probe streams the same short prompt to every model through the same clients the site itself uses, records time-to-first-token and total wall time, and repeats it 3 times. The table shows medians. Probe spend lands in the site's own AI-spend ledger like any other call.
Throughput, honestly
Throughput is output tokens per second over the generation phase only (total time minus TTFT): the stream speed a reader experiences. DeepSeek R1 streams visible chain-of-thought, so its first token is honestly slow; the row says so instead of hiding it.
Prices only where they exist
Anthropic rows show committed vendor list prices, imported from the same table /benchmarks uses so the two can never drift. Open models route through the HF Router, which bills pass-through at whichever provider it picks per request, so those cells honestly read n/a instead of an invented number.
Runner writes, page renders
This page never calls a model. A hand-run probe (npm run inference:probe) writes raw observations into the repo; the page renders them, your browser re-derives every number from the raw trials, and CI fails if the stored summary ever disagrees.
Why measure like a provider?
Because publishing a provider table is step one of becoming a provider. Hugging Face documents exactly how a provider joins its table: an OpenAI-compatible API, registered model mappings, a billing endpoint, and automated probes every six hours (the register-as-a-provider guide ↗), and OpenRouter publishes its own bar for listed providers (for-providers ↗). Those tables run on the same three instruments this page just built: a measured roster, honest prices, and probes with receipts. The serving side comes later; the measuring side now works.
Keep exploring
The same models graded on accuracy, speed and cost, with live self-checking scores.
Hugging Face field notes →The expedition report on the open-model commons this table borrows its shape from.
Chat →Talk to these models yourself; the picker offers the same roster this table measures.
The Council →Five of these models answering in parallel, reconciled by a sixth.
Watch the table go live
An occasional note when the probe runs for real, a model joins the roster, or the serving side of the provider arc ships. No schedule, no filler.