Methodology
A sourced value has a URL, a short quote, and the date it was read. If a price was not on the page, the field is empty and marked “not published / unverified”. A value whose last-checked date is more than 30 days old is flagged stale. Every platform starts unverified. Verification is a separate flag, set only by an approved correction.
Cost per 1,000 runs
C_1k = 1000 * Σ_s (t_in_s*(1−h_s)*p_in + t_in_s*h_s*p_cached + t_out_s*p_out) * (1+r_s)
Token counts use the preset low, point, and high. Cache hit rate and retry rate stay point assumptions. If a cached price is not published, the cache-hit term is dropped and those tokens are billed as ordinary input. The chart shows the range. The point is the preset, not a measurement.
Self-hosted and dedicated GPUs
GPU-hours to serve peak and average output tokens, divided by runs, plus an operations overhead. The default overhead is 15% and the compare page lets you move it. Per-second billing uses busy GPU time only. Provisioned billing keeps an integer GPU count for the peak window and another for the rest of the day. A published per-second rate is multiplied by 3600 inside the cost engine, and that conversion is labeled. Node-hour prices are shown and left out of C_1k.
Tokens per second is your measurement. The catalog does not invent one, so a GPU row has no run cost until you type it.
Quality
A point is plotted only when the preset names a public benchmark and the catalog has a cited score for that offering. This snapshot has no such scores, so compare says “no eval yet” and does not draw a frontier over the live rows.
Cost versus quality
Example geometry. These letters are not model scores.
The letters above are a geometry check: B, C, and D sit on the frontier, and A is inside it. They are not evaluations of any model.
Share links and versions
Each catalog version is an immutable snapshot. A compare link stores that version plus the preset inputs, so opening it later replays the same inputs against the same prices.