Best buys: the quickest pay-back at each level of capability

For each amount of daily use, the machines and models that pay for themselves soonest, grouped by how capable the model is. Each class lists the three that pay back soonest, one row per model, on the machine that pays it back quickest; under each table, a class counts what it leaves out and names the next model down. For every model on the site with its class beside it, see the leaderboard.

The short version: at 1M tokens a day (a moderate coding-assistant day), the quickest Sonnet-class pay-back is Qwen3.8 27B on a Mac mini M6, 32GB, in 41 years.

Jump to: 50k/day · 200k/day · 1M/day · 4M/day · 20M/day

Current machines at list price and current models only. Graphics cards are priced as the card alone, so add the PC around it before comparing them with a complete computer; the cards are ranked against each other here. Every row uses 15:1 input to output, $0.17 per kWh, 32k of context, and today's API prices held flat; switch on falling API prices in the calculator and the years stretch. Speeds marked estimated are worked out from memory bandwidth rather than measured, and pay-back scales with them. Models nobody rents are priced as their closest hosted match and say so. Scores are the Artificial Analysis Intelligence Index v4.3; * marks a score the index estimated. Open any row to change the assumptions, set two machines or two models against each other in the head-to-heads, or read what a million tokens costs to rent against generating it, which is the gap every figure here divides into.