OpenRouter Pricing Explained: Per-Token Costs

A million tokens of Claude Sonnet 5 costs $2 going in and $10 coming out. A million tokens of Nex-N2-Mini costs two and a half cents. Same platform, same API call, a 200x price gap — and if that sentence made complete sense to you, you can skip this article. For everyone else, OpenRouter pricing is the clearest window into how the entire AI industry bills for intelligence, because the platform lists hundreds of models side by side with their raw per-token rates. This guide explains what those numbers mean, why they vary so wildly, where the hidden fees sit, and how to know when per-token billing is the wrong deal for you altogether.

How Does OpenRouter Pricing Actually Work?

Every request has two meters running. Input tokens (also called prompt tokens) are everything you send: your question, the system prompt, the conversation history, that PDF you attached. Output tokens (completion tokens) are everything the model writes back. Each is billed at its own rate, quoted in dollars per million tokens, and a token is roughly three-quarters of an English word — so a million tokens is a decent-sized bookshelf.

Output always costs more, usually 3-6x more. Generating text is more computationally expensive than reading it, and pricing follows the electricity. Kimi K3 reads for $3 per million and writes for $15; Grok 4.5 reads for $2 and writes for $6. This asymmetry matters for budgeting: a chatbot that gives long answers spends most of its money on output, while a document-analysis tool that reads huge files and returns short verdicts lives on the input meter.

Reasoning models add a twist — their internal "thinking" is billed as output too, even though you may never read it. Point a heavy reasoner at a simple task and you can pay for ten paragraphs of deliberation behind a one-line answer.

What Are Cache Reads, and Why Are They 90% Cheaper?

Look closely at any model listing and you will find a third number: the cache-read rate. When your requests keep re-sending the same prefix — a long system prompt, a fat document, yesterday's conversation — providers can store that prefix after processing it once, then charge a fraction to "re-read" it from cache. The discounts are steep: Claude Sonnet 5 charges $0.20 per million cached input tokens against $2.00 fresh; GPT-5.6 Luna drops from $1.00 to $0.10. That is a 90% discount for being repetitive.

For agentic workloads that loop over the same context hundreds of times, caching is the difference between a reasonable bill and a horrifying one. It is also a good example of why per-token pricing rewards engineering attention: the developers who structure prompts for cache hits pay a fraction of what naive implementations pay for identical results.

Real Per-Token Prices in July 2026

These are live rates, per million tokens, for a representative spread of the catalog (the same models are browsable in CoreAI's model library):

Note that these numbers move. Sonnet 5's $2/$10 is introductory — Anthropic's official pricing page lists $3/$15 as the standard rate taking over after August 31, 2026. A 50% quiet increase on the most popular coding model is exactly the kind of thing per-token budgets have to absorb; our Sonnet 5 pricing breakdown has the details.

Why Do Prices Vary So Much Between Models?

Three forces set the spread. Compute: bigger models on more expensive hardware cost more to run — Kimi K3 has 2.8 trillion parameters and prices accordingly, while Nex-N2-Mini is small enough to serve for pocket change. Positioning: premium flagships like Fable 5 and GPT-5.6 Sol price for the customers who need the last 5% of capability, the same way business class subsidizes the rest of the plane. Competition: open-weight models are hosted by multiple providers bidding against each other, which is why the GLM, Qwen, and Kimi families keep drifting cheaper while closed flagships hold their prices.

On top of the per-token rates sits the platform's own fee structure. Token prices carry no markup, but buying credits costs 5.5% (minimum $0.80, or a flat 5% via crypto), and bring-your-own-key traffic is free for the first million requests a month, then 5% of the equivalent cost. Fair, transparent — and one more line item to track.

How Do You Keep Per-Token Costs Down?

If you do live on the meter, four habits cover most of the savings. Route by task: send mechanical work to the budget tier and reserve flagships for the output someone will actually judge — the price ladder in our budget models guide exists to be climbed deliberately. Trim the context: conversation history re-sends with every message, so a 40-message thread is quietly billing you for 39 messages of archaeology each turn; start fresh threads more often. Structure for cache hits: keep your stable instructions in an unchanging prefix so the 90% cached rate applies. And cap your outputs: asking for "a concise answer" is not just style advice when the output meter runs at five times the input rate. None of this is hard — but noticing that it is work is the point. The meter only stays cheap if someone keeps watching it.

What Does a Real Month Actually Cost?

Concrete scenario: a heavy individual user who sends 300 messages a day, averaging 2,000 input tokens (with history) and 500 output tokens each. That is 18M input and 4.5M output tokens a month. On GLM 5.2: about $30. On Claude Sonnet 5: $81. On GPT-5.6 Sol: $225. Same person, same conversations — the only variable is the model, and the bill swings by 7x. Add the top-up fee, the cache-hit lottery, and a couple of long-document days, and you see the real problem: the meter is efficient, but it is never predictable. Developers accept that trade because they can optimize it. Most people cannot, and should not have to.

That predictability gap is exactly what flat-subscription apps sell. CoreAI gives you 300+ models — including every model in the table above — for one fixed monthly price, with usage limits handled for you instead of a running dollar meter. You pick models by quality, not by fear of the invoice, and you can sanity-check any choice by running two models head-to-head in Compare. If your monthly AI usage looks like "a lot, but I have no idea how many tokens," a flat plan is almost certainly the better deal — the same logic covered in our subscription consolidation piece.

Frequently Asked Questions

Does OpenRouter add a markup to model prices?

No — per-token rates are the providers' list prices, unchanged. The platform earns its keep through a 5.5% fee (minimum $0.80) when you purchase credits, plus a 5% fee on bring-your-own-key requests beyond the first million each month.

What is the difference between prompt and completion pricing?

Prompt (input) tokens are everything you send the model; completion (output) tokens are everything it generates, including any hidden reasoning. Output rates run 3-6x higher than input because generation costs more compute than reading.

How much does a typical AI conversation cost per token?

A single exchange — a 200-word question, a 400-word answer — costs a fraction of a cent on budget models and a cent or two on flagships. Costs compound through conversation history: every new message re-sends the whole thread as input, which is why long chats quietly get expensive and why cache pricing exists.

Why did my per-token bill jump with a reasoning model?

Reasoning models bill their internal thinking as output tokens. A hard problem can generate thousands of deliberation tokens you never see. Use reasoning modes where they pay off — our thinking-mode guide maps when that is — and a cheaper direct model everywhere else.

Is a flat subscription cheaper than per-token billing?

For developers running optimized batch workloads, per-token usually wins. For humans having conversations, a flat plan wins on predictability alone: heavy chat on a frontier model can meter past $80 a month without you noticing, while a fixed plan costs the same in a busy month as a quiet one.