LLM price comparison

Every major model's API pricing in one live table — sort by input or output cost, filter by provider, and see what a request of your size costs on each. Prices load fresh from OpenRouter every time, so the table never goes stale.

How LLM API pricing works

LLM APIs bill per token, and every model carries two rates: one for input (your system prompt, question, retrieved documents, and conversation history) and one for output (the response the model generates). Both are quoted per million tokens, so the numbers in the table look small — a request is typically a fraction of a cent. Output is almost always billed at a higher rate than input, often several times more, because generating text costs more compute than reading it. When you read a row, look at both columns: the input rate governs how expensive it is to send context, and the output rate governs how expensive it is to get a long answer back.

How to actually compare models on price

Comparing headline rates alone is misleading. A model with the lowest input price can still be the most expensive choice for a workload that returns long answers, because the output rate does the damage. The honest comparison weights each rate by yourinput-to-output ratio and your request volume. A retrieval-heavy RAG chatbot sends thousands of context tokens per call and writes short answers, so it's dominated by the input rate; a brainstorming or drafting tool writes long responses, so it lives or dies on output. Set your typical token counts above and rank models on the per-request cost, then multiply by daily volume — that is the number that decides the bill, not the headline rate.

Discounts that change the real price

The list price isn't always what you pay. Prompt cachingstores a repeated prefix — a long, fixed system prompt or knowledge base — so you aren't charged full input price for it on every call; providers discount cached reads steeply, which is a large saving for RAG and agent workloads that resend the same context each request. A batch API runs asynchronous, non-time-sensitive jobs at roughly half price, making it the right home for backfills, evaluations, and bulk generation. High-volume accounts can also reach provider volume tiers with lower committed rates. None of these show in a headline table, so factor them in when your workload has repeated context or tolerates latency.

Price versus quality

The cheapest model that passes your evaluation wins — not the cheapest model overall. A model that costs a tenth as much but needs a retry, or produces output you have to throw away, isn't actually cheaper once you count the wasted calls and the human time to fix it. The durable pattern is to right-size per step: use a small, fast model for routing, classification, and extraction, and reserve a frontier model for the one step that genuinely needs its reasoning. Price is a tie-breaker among models that clear your quality bar, so build the eval first and let it decide which rows in the table you're even allowed to compare.

Frequently asked questions

How is LLM API pricing calculated?
APIs bill per token, not per request or per word. Each model has two rates — one for input tokens (your system prompt, question, documents, and history) and one for output tokens (the response) — quoted per million tokens. Your cost for a call is input tokens × input rate plus output tokens × output rate. Because the rates are per million, a single request is usually a fraction of a cent; the bill grows with request volume.
Why is output more expensive than input?
Generating tokens is more compute-intensive than reading them — the model runs a full forward pass for every token it writes, whereas input can be processed in parallel. Providers price this in, so output typically costs several times more than input on the same model. That means response length, not just prompt size, drives spend on chatty workloads.
Which LLM API is cheapest?
There is no single cheapest model — it depends on your input-to-output ratio and volume, and prices change often. Sort the live table by input or output rate to see the current order, but the model with the lowest headline input price isn't automatically cheapest for you: a model with a low input rate but a high output rate can lose to a pricier-looking one on a workload that writes long answers.
How do I compare model prices fairly?
Don't compare headline rates in isolation. Estimate your typical input and output token counts, then price a real request on each model — the per-request column does this so you can rank models on what your workload actually costs. Weight by volume too: a difference of a hundredth of a cent per call is meaningless at 100 requests a day and thousands of dollars a month at scale.
What is prompt caching and batch pricing?
Both are discounts that change the real price. Prompt caching stores a repeated prefix — a long system prompt or fixed knowledge base — so you aren't billed full input price for it on every call; providers discount cached reads heavily, which helps RAG and agent workloads that resend the same context. Batch APIs run asynchronous, non-urgent jobs at roughly half price, ideal for backfills, evals, and bulk generation. Large-volume accounts may also negotiate lower tiers.
Does the cheapest model always win?
No — the cheapest model that passes your evaluation wins. A model that's a tenth of the price but needs two attempts, or produces answers you have to discard, isn't actually cheaper. Use a small, cheap model for simple steps like routing, classification, and extraction, and reserve a frontier model for the one step that genuinely needs it.
Is this comparison free?
Yes. The table is free, needs no signup, and loads current prices live so it never goes stale. Set your own token counts to cost out your workload, then estimate monthly totals in the linked cost calculator.

Related tools: LLM cost calculator (per-request, daily, and monthly spend), RAG cost calculator (embeddings + storage + generation), and the context window comparison (count tokens and check what fits).

More free tools

  • LLM cost calculator — Estimate per-request, daily, and monthly spend from token counts and request volume.
  • RAG cost calculator — Full pipeline costs — embeddings, vector storage, and generation — from corpus size and query volume.
  • RAG chunking visualizer — Paste a document and watch chunk size, overlap, and strategy change the chunks before you embed.
  • LLM context window comparison — Count the tokens in any text and check which models it fits, against live context windows.
  • Tool call & MCP schema builder — Build function-calling parameters once, copy valid JSON Schema for OpenAI, Anthropic, and MCP.
  • AI chatbot ROI calculator — Turn ticket volume, handle time, and deflection rate into monthly support savings and ROI.