What a million tokens costs.

Input and output are priced separately, and output is where the bill actually comes from.

Output tokens cost 3–6× input, which inverts most cost intuitions

Almost every provider prices output several times higher than input, so the model that looks cheapest on the headline input number is frequently not the cheapest in production. What decides your bill is the ratio of tokens you send to tokens you get back. A retrieval-augmented app sends enormous context and receives a paragraph, so it is input-dominated and should optimise for input price. A chat product sends a short prompt and streams a long answer, so it is output-dominated and the same comparison reverses. Prompt caching moves the line again by repricing repeated prefixes at roughly a tenth of input rate — which is why a cached-prefix agent can be dramatically cheaper than its sticker price implies.

Qwen3.7 Flash
$0.03 in · $0.13 out — 1M context
DeepSeek V4 Flash
$0.08 in · $0.25 out — 1M context
GPT-5.6 Luna
$0.10 in · $0.60 out — 1.05M context
Gemini 3.5 Flash Lite
$0.30 in · $2.50 out — 1M context
GLM 5.2
$0.76 in · $2.42 out — 1M context
GPT-5.6 Terra
$1.00 in · $6.00 out — 1.05M context
Gemini 3.6 Flash
$1.50 in · $7.50 out — 1M context
Claude Sonnet 5
$2.00 in · $10.00 out — 1M context
Claude Opus 5
$5.00 in · $25.00 out — 1M context
GPT-5.6 Sol
$5.00 in · $30.00 out — 1.05M context

Figures verified 2026-08-10 — OpenRouter model API, checked 2026-08-10. Per 1M tokens, USD.. Model pricing moves; treat that date as an expiry and re-check before deciding anything on it.

FAQ

Why is output more expensive than input?

Input is processed in parallel across the whole prompt in a single forward pass, while output is generated one token at a time, each pass depending on the last. Output is inherently serial and therefore occupies the hardware for longer per token. The pricing reflects the compute shape, not a margin decision.

Which model is cheapest overall?

That depends entirely on your input:output ratio, which is why a single ranking is misleading. Compute your own: multiply your average input tokens by the input price and your average output tokens by the output price, per request, then multiply by volume. The ordering frequently differs from the headline-price ordering.

How much does prompt caching change this?

A great deal for anything with a stable prefix — a long system prompt, a fixed tool set, a document being asked about repeatedly. Cached reads are typically around a tenth of the input rate, so an application that reuses a large prefix across many calls can see most of its input cost disappear. It does nothing for output.

Are these numbers current?

They were checked on 2026-08-10 against the OpenRouter model API. Model pricing moves — sometimes several times a quarter — so treat that date as an expiry. If it is more than a couple of months old, verify before making a decision on it.

Related