How Much Does the OpenRouter API Cost? GPT, Claude, and Gemini per Million Tokens and the Monthly Bill

How OpenRouter charges: no monthly fee, then a fee on credits

The $2 / $10 on a model page is not yet the card charge. Inference is not marked up. The OpenRouter FAQ says OpenRouter takes the provider’s token count and deducts that provider’s list price from your credits. The platform fee is charged when you buy the credits.

The free plan on the OpenRouter pricing page is 50 requests a day and has no payment method. Standard and Business are both pay as you go, not a seat subscription. The fee percentages differ, and so do the features: budgets, routing, and prompt caching start on Standard; in-region routing is on Business. Enterprise is listed as negotiated fees and invoicing.

• The 5.5% comes out of the top-up. It is not a 1.055 multiplier on the token price. To land credits equal to an inference total C, when C is at least about $13.75, the card charge is C / 0.945.
• Below that, the $0.80 minimum beats 5.5%. $0.80 is 5.5% of about $14.55 charged, which leaves about $13.75 of credits.
• Unused credits can be refunded for 24 hours after purchase. The platform fee is not refunded, and crypto is not refundable. The terms reserve the right to expire unused credits one year after purchase. Balance is the credits endpoint endpoint.

What GPT, Claude, and Gemini cost per million tokens

The table is the standard list price read from OpenRouter models API on 2026-10-13, in US dollars per million tokens. It skips :batch rows and -pro variants. prompt is input, completion is output, and input_cache_read is the cache-read price. Treat the live model page and that endpoint as newer than this table. It is not a contract rate.

On the same endpoint, :batch is about half the standard price. claude-sonnet-5.5:batch is $1.00 input and $5.00 output. gpt-6-sol and gpt-6.1-sol are both $2 / $10 on input and output; cache reads are $0.20 and $0.10, so a bill with no cache is the same. On the GPT-6 Sol model page page, the OpenAI endpoint used in standard routing is $2 / $10. Endpoints marked off standard routing include Azure around $2.20 / $11 and OpenAI Fast at $4 / $20. A default request does not land on Fast.

How the monthly bill is calculated

This is price math, not a readout from our account. Neither case includes cache, images, or web search. The light month is 10 million input tokens and 1 million output tokens. The agent month is 50 million input and 5 million output. Inference = millions of input × input price + millions of output × output price. The card column uses the OpenRouter FAQ card rule and assumes one top-up that exactly covers that month: at or above about $13.75 of inference, charge = inference / 0.945; otherwise charge = inference + $0.80. An existing balance, or several smaller top-ups, changes the fee. gpt-6.1-sol matches gpt-6-sol on this table because cache is excluded.

The agent month on Sol or Sonnet alone: $150.00 of inference, $158.73 on a card, of which $8.73 is the platform fee. Luna or Haiku in the light month is only $1.50 of inference but $2.30 charged, because the $0.80 floor is larger than 5.5%. If a 5% crypto fee is also taken out of the top-up, the agent month’s $150 costs about $157.89 (150 / 0.95). Business at 8% is not in the table. The pricing page does not publish the same formula as the 5.5% rule; if 8% is also taken from the top-up, the charge is about $163.04 (150 / 0.92). If the contract uses another base, use the invoice.

Where cache, batch, and routing move the same bill

Claude Sonnet 5.5 cache reads are $0.10 per million tokens. Standard input is $2. If 80% of the agent month’s 50 million input tokens are cache reads: 10 million × $2 + 40 million × $0.10 + 5 million × $10 = $74. With no cache the same month is $150. The $76 gap is the assumption, not a production log. Cache helps when the system prompt and tool definitions repeat.

:batch on the same model is half the standard list price. If the whole agent month uses claude-sonnet-5.5:batch at $1 / $5: 50 million × $1 + 5 million × $5 = $75. Batch buys waiting time. It does not turn a flagship into a cheap tier. Check that variant’s own rate limits and queue before dividing the price by two in a launch plan.

In the OpenRouter FAQ, :floor sorts providers by price and :nitro sorts them by throughput. :floor does not turn GPT-6 Astra ($10 / $50) into GPT-6 Luna ($0.10 / $0.50). The first cut is the tier. On the same agent month, Luna or Haiku is $7.50 of inference and Astra is $750.00. The card totals are $8.30 and $793.65. The tier gap is on the order of fifty times. The 5.5% fee is not.

When the provider’s own API is the better bill

OpenRouter’s margin is not on the token price. It is on the credit purchase, in exchange for one key across providers and a fallback when an endpoint errors. On the agent month, Sol or Sonnet inference is $150 and the card adds about $8.73. If you already have the three provider accounts, a contracted rate, or you need Enterprise invoicing, that $8.73 does not buy the missing piece. If you want one bill and automatic failover, that $8.73 is the price of it.

BYOK waives the extra 5% on the first $25,000 per month of list-price inference. Above that, 5% of the OpenRouter price comes out of credits, and the provider still bills you. Google’s own free tier and per-tier Gemini rates are a different table: Gemini API Pricing 2026: Token Cost and Free Tier Explained. A Claude Code seat is not this token table either: How Much Does Claude Code Cost Per Month? 2026 Pricing & Plans Explained. Do not add a seat fee and an OpenRouter balance and call the sum an OpenRouter monthly plan.

• OpenRouter says it does not offer volume discounts. At high volume, a provider contract can sit under these list prices.
• Prompts and completions are not logged by default. Opting in is a 1% usage discount in the FAQ. That is a discount, not a second price list.
• There are 25+ free models, but 50 requests a day (1,000 after 10 credits) is not a production path.

FAQ

Does the OpenRouter API have a monthly fee?

There is no monthly plan. Standard and Business are pay as you go: buy credits, then each request deducts tokens. The platform fee is on the credit purchase, not a seat.

Is the per-million price the same as the provider’s?

The FAQ says inference is not marked up. You pay the list price of the provider that served that request. Two providers can list different numbers for one model. This table is the 2026-10-13 catalog price and will change.

Why can a small top-up look like more than 5.5%?

Card purchases have an $0.80 minimum. 5.5% only reaches $0.80 at about $14.55 charged. Under about $13.75 of inference, one exact top-up costs the inference total plus $0.80, so the percentage is higher than 5.5%.

Is the Business 8% already in the monthly table?

No. The card column uses only the FAQ’s 5.5% card fee and the $0.80 floor. The pricing page lists Business at 8% without the same formula. If 8% is also taken from the top-up, divide by 0.92. If the contract uses another base, use the invoice.