2264 model families, 187 providers, 4460 price changes in 30 days

Prices are re-read from every provider daily and appended to a ledger, never overwritten. 2364 of the recent changes were drops. Watch a family to get an email the morning after it moves.

Price changes, last 30 days

Showing the 40 most recent of 4460.

Price index, last 90 days

Median list price of every paid chat listing the catalog saw that day, one price per seller, in USD per million tokens. Only sources that polled that day count, so a missing day is a gap rather than a guess. The index is not held to a fixed basket: when a new source adds listings the median moves with them, and the listing count below says when that happened.

2026-11-13: 5,165 listings from 7 of 7 sources; cheapest paid input $0.007, output $0.010. Over 90 indexed days, 951 listing-cuts and 861 listing-raises; 44 days had a source missing. Series by provider, family and /decide task via /api/v1/trends.

Price catalog

One row per model family, ranked by how many providers compete on it. Cheapest is by input + output list price; free and preview tiers are excluded. Full list at /api/v1/families.

Showing the 60 most-contested of 2264 families.

2000 throughput numbers verified by ≥2 independent measurements

Every cost quote on this site sits on numbers that came from somewhere real. We don't average estimates and call it a benchmark. When two independent labs (NVIDIA, Dell, Cisco, Lenovo, AMD, ASUSTeK, Google, Oracle, Supermicro, GigaComputing) publish the same MLPerf result within 10%, we mark itverified-cross-checkedand surface it here.

Best deploy: Llama 3.1 70B, 1.2M tokens

The end-to-end optimizer pulls hosted-API pricing, GPU rental hourlies, and self-deploy electricity (with PUE and grid carbon) into one ranked list. Same workload, three honest answers: cheapest dollars, greenest emissions, and the closed-source alternative.

Full ranked list: POST /api/v1/cost/best-deploy with {family, inputTokens, outputTokens}. Returns hosted-API + hosted-GPU + self-deploy options ranked by cost; closed-source families return hosted-API only.

Headline agreements

Sorted by number of independent labs that published the same benchmark. Delta is the tightest pairwise agreement. Click through to MLPerf v5.1 for the raw submissions.

Cheapest self-host options

Cost per million tokens for cross-checked throughput rows joined against verified public hourly GPU pricing. Single-node configurations only (multi-node MLPerf rows excluded — not rentable at public per-hour rates). Math:($/hr × gpu_count × 1M) / (tok/s × 3600)

Electricity by major AI city

Industrial meter rate × PUE (cooling overhead) = what the GPU actually pays for power. Same hyperscale workload (8×H100 SXM running MLPerf Llama-2-70B at 5,500 tok/s, 5,600W IT draw, PUE-adjusted facility draw). 5.3× differential from cheapest to most expensive. Sources: EIA (US, verified), Eurostat (EU + UK, verified), DESNZ (UK direct, verified), training approximations marked.

Pass ?cityCode=us-memphis to /api/v1/cost/run-locally and the calculator uses that city's rate instead of the $0.12 default. Full roster: /api/v1/electricity.

How cross-check works

Full JSON roster: GET /api/v1/health/cross-check. Filter by ?axis= or ?family=. Each row carries the full evidence used in the upgrade.