Pareto pricing: calculate the token-mix crossover correctly

Published 2026-09-30 · updated 2026-11-18

On November 18, 2026, Cloudflare’s Pareto listing describes Unbiased’s blended model and quotes 2.50 USD per million input tokens, 0.25 USD per million cached input tokens and 7.50 USD per million output tokens. These are that provider’s listed terms, not a universal quote across all hosts.

For an illustrative 6 million uncached input tokens and 1 million output tokens, the listed token rates give 6 × 2.50 + 1 × 7.50 = 22.50 USD. Confirm applicable account terms and any other charges before budgeting. This article does not measure quality or claim Pareto sits on a performance frontier.

Repairing the earlier crossover formula

The previous article asserted a 20 percent input discount and 15 percent output premium over another model without establishing a dated provider pair. Treat those percentages as hypothetical assumptions, not today’s Pareto comparison.

Let the reference input rate be a, its output rate be k × a, and the fraction of all tokens that are output be s. The alternative’s rates are 0.8a and 1.15ka. Costs per total token, in units of a, are:

reference   = (1 - s) + k × s
alternative = 0.8 × (1 - s) + 1.15 × k × s

Equating them gives 0.20 × (1 - s) = 0.15 × k × s, hence:

s = 0.20 / (0.20 + 0.15 × k)

At k = 4, the crossover is 25 percent output. Below that share, the alternative is cheaper under these assumptions; above it, the reference is cheaper. The former expression 1/(1+4/k) was wrong.

At a 6:1 input-to-output ratio, reference cost is 6 + 4 = 10 units and alternative cost is 6 × 0.8 + 4 × 1.15 = 9.4, a 6 percent saving. At 1:3, the costs are 13 and 14.6 units: about 12.31 percent more for the alternative.

Apply the method to actual quotes

Get both input and output rates for each exact model/provider pair from the model catalog and verify them at the provider. Never combine the cheapest input half of one offer with the cheapest output half of another.

Use actual billed categories, including cache and reasoning where applicable. The simple formula assumes uncached text, constant rates and no fixed fees. Test quality and retry rates separately: a lower token bill does not establish lower cost per accepted task.

Use the calculator for scenario checks and the cost worksheet for a fuller comparison. The corrected algebra is reusable; the hypothetical discount percentages are not a live market claim.

Put the numbers to work. Run your own workload through the breakeven calculator with today's floors, then price hardware: rent at Vast.ai → · rent at RunPod → Rental links are referral links. They never affect our rankings, which are ordered by price alone.

← All guides