Compare configurations within each model using the pages in this section. Different models do not deliver interchangeable output or quality.
Read the cost curve
Each solid curve shows recorded aggregate output throughput and GPU cost per million output tokens for one setup and recipe. Both axes use linear scales starting at zero. The cost axis ends at three times the highest minimum cost across the displayed curves, rounded up to a readable tick. High-cost samples outside this range remain in the source snapshot. A compact table compares providers, GPUs, and cost per 1M output tokens at peak measured throughput. Connected samples describe a benchmark run, not a guarantee of throughput at an arbitrary load.
Your effective cost depends on utilization. Idle GPUs still incur hourly charges. These figures do not include all possible storage, network, input-token, or operational costs. Iwant is free; cloud resources can incur charges.
Measured and estimated costs
A cloud price estimate must use the exact same accelerator configuration. Scale measured cost by alternative hourly price divided by measured hourly price, leaving throughput unchanged. This assumes equal throughput across clouds. Missing prices and zero measured hourly prices cannot support an estimate. A price does not guarantee capacity.
Dashed curves compare the cheapest listed on-demand region per cloud and exact GPU configuration. Models without eligible offers show measured curves only. Hosted completion-price medians appear separately because their workload and input-cost accounting have not been verified.
Reproduce and refresh
Model pages record recipe versions, source run timestamps, and original launch commands. Original commands may differ from the current CLI. Follow the launch guide for current options.
Run node scripts/iwant/sync.mjs in the documentation repository to refresh committed snapshots and pages from the public benchmark API. Review the generated changes before publishing. Run timestamps are not GPU price snapshot dates. The source does not currently establish price freshness. Last modified on October 31, 2026
