LLM Pricing Comparison

On sticker price, DeepSeek, at roughly $0.28 per million input tokens. On your actual bill, it depends on your traffic shape: output tokens bill 4x to 9x input, and cache discounts range from 50% to 90% between providers, so the cheapest model at one input/output mix is regularly not the cheapest at another. The comparison above ranks by effective monthly cost at your own numbers.

Three reasons. Output multiples differ, so output-heavy work reorders the table. Cache discounts differ, so a stable prefix at high traffic can make a dearer model cheaper in practice. And failure rates differ: a cheap model that retries pays for every failed attempt, which is invisible in a price table and dominant in a real bill.

Price a month of your real traffic, not a million tokens. Take your input and output tokens per request, your request volume, and your cache hit rate, and compute the monthly figure per model. Then check quality on your own evaluation set, because the cheapest model that fails your task is the most expensive option on the list.

Inference cost is price times volume across three token types: input, output at 4x to 9x the input rate, and on reasoning models the internal thinking tokens, which bill as output without appearing in the response. The comparison above prices your mix directly; the per-token rate alone is the least informative of the three numbers.

Several times a year per provider, usually downward, and occasionally with structural changes like new cache tiers that move effective cost more than the sticker change does. This page shows its verification date and links each provider directly, so you can confirm before committing an architecture to a number.

Rarely as a first move. Provider choice spans roughly 2x on a comparable tier, while fixing caching and routing on the provider you already use spans 3x or more, without a migration or re-evaluating every prompt against a new model. Switch for capability, latency or rate limits; optimise for cost.