What a million tokens actually costs.

• Rates are read live from each provider’s own API — refreshed on this page every three minutes, cached ten minutes at the edge. Last read seconds ago.
• MEAS is a speed we measured on that card with that model, from a rig report. EST is our engine’s prediction — a cost built on it inherits the engine’s error.
• Cost per million tokens is the hourly rate divided by output tokens at one stream. Batched serving is cheaper per token; an idle card still bills.
• Datacenter cards rank badly here, and that is the finding, not a fault. An H100 or H200 earns its price by serving many streams at once; at one stream its enormous bandwidth sits mostly idle, and our engine models that as a much lower efficiency than a consumer card gets. If you are serving a crowd rather than yourself, this ranking inverts — that case is the enterprise planner, not this page. Those rows are also the least certain on the page: they are estimates, and the constant behind them is the one we most want to replace with a measurement.
• A marketplace price is one host’s listing, not a published rate: it can vanish, and hosts may run a card below its stock power limit, which costs speed the price does not show.
• Only cards holding the model at the chosen quantisation, with 10% left for the runtime, are listed. Nothing here is priced through CPU offload.

Provider links may carry a referral code. It costs you nothing, it does not change what is listed, and it never changes the order — the table sorts by the column you choose and by nothing else. See methodology.