OpenRouter pricing, explained

What OpenRouter charges

Unused credits can be refunded within 24 hours of purchase, minus the fee, and may expire a year after purchase. Check OpenRouter’s FAQ for changes.

What tokens cost on OpenRouter today

The newest models from each major provider, cheapest first. Prices are dollars per million tokens. The LLM cost calculator has all 448 models and turns them into a monthly bill.

A worked example

Say a support summarizer runs 5,000 times a day, sending 3,000 tokens and getting 500 back.

• On Claude Sonnet 5.5 ($2.00 in, $10.00 out per 1M), one call is 3,000 × input + 500 × output. Run that 150,000 times a month and the fee on top is 5.5% of whatever credits you buy to cover it.
• On DeepSeek V4.1 Flash ($0.30 in, $1.20 out), the same month costs a fraction of that, if the summaries are still good.

The model choice moves the bill by 10× or more. The 5.5% fee moves it by 5.5%. Run these numbers in the calculator.

Caching discounts on OpenRouter

OpenRouter passes through each provider’s prompt-cache pricing. Repeated prompt prefixes are billed at the cache-read price, often 10–25% of the input price. Anthropic and Gemini need explicit cache_control breakpoints; OpenAI, DeepSeek, Grok and most open models cache automatically. OpenRouter keeps you on the same provider for a while after a cache hit, and a session_id makes that stickiness reliable. Details and numbers: prompt caching costs by provider.

What OpenRouter’s dashboard can’t tell you

OpenRouter’s activity page shows spend by key, model and app. It can’t tell you that the nightly digest costs $40 a day while chat costs $3, because it never sees which feature made the call. That is the number you need before switching models: you only want to move the features that are expensive and easy.

Options, from most work to least:

1. One OpenRouter key per feature, then compare keys on the activity page.
2. Log usage from every response with a feature tag into your own analytics.
3. Wrap your client with Gin: fetch: gin({ useCase: 'digest' }). You get spend by use case in a free daily email, and cheaper models graded on your real calls.

How to lower an OpenRouter bill

1. Move the easy features first. Classification, extraction, short summaries and follow-up turns usually pass on a model 5–20× cheaper. Check on your own prompts before you switch, not on benchmarks.
2. Make prompts cacheable. Static instructions and tools first, the per-request part last, and a stable session_id per conversation or feature.
3. Cap output. Set max_tokens, ask for terse JSON, and turn reasoning off or down where it doesn’t help.
4. Use batch variants for offline work. Models with a :batch suffix on OpenRouter are listed at about half price.
5. Keep a fallback. Any cheaper model will fail some prompts. Route those back to the original model instead of shipping worse answers.

Gin automates 1, 2 and 5: it replays your recent calls on cheaper candidates, has a judge compare each answer with the original, and only routes a use case once ≥90% of its replays came back the same or better. Everything else stays on your model, on your OpenRouter key.