LLM API pricing, 2026

Every large-language-model API is priced the same way — a rate per million input tokens and a separate, higher rate per million output tokens — and yet the bill for the same workload can differ by 20x between providers. This page tracks the current rates and links to head-to-head comparisons for the pairings people actually weigh up.

The three numbers that decide your bill

Input price is what you pay for everything you send: the system prompt, the conversation history, any retrieved documents, the user's message. For most real applications this is the larger half of the bill, because context grows with every turn and RAG pipelines stuff thousands of tokens of retrieved text into each call.

Output price is what you pay for what the model generates, and it is always higher — typically four to eight times the input rate. A workload that produces long answers, writes code, or drafts documents is output-heavy and should be priced on that column.

Cached-input price is the one people forget. If you send the same long prefix — a fixed system prompt, a static knowledge base, a few-shot block — most providers will cache it and bill the repeated portion at a fraction of the normal input rate, often a tenth. An agent that replays a 4,000-token instruction set on every step lives or dies on this number.

How to actually compare two providers

Take a representative request from your app. Count the input tokens and the output tokens (roughly: one token per 0.75 words of English). Multiply each by the provider's rate, add them, and multiply by your monthly call volume. The [cost calculator](/calculator) does exactly this across every model at once — enter your volume and it ranks the bill. Do not compare on the headline "flagship model" price if your workload runs fine on a mid-tier model — the gap between a frontier model and its smaller sibling is often 15x for a quality difference most apps never notice.

Where the rates come from

Each figure on the individual comparison pages is checked against the provider's own pricing page or its published model rates, cross-referenced with OpenRouter's live model API, and stamped with the date it was verified. When a provider changes its pricing, the affected comparisons are updated and the change is logged. The full dataset — every provider, every model, machine-readable — is on the [data page](/data).