OpenRouter vs Direct API Pricing: Are You Paying a Middleman Tax?

OpenRouter's pitch is one API key for hundreds of models, and its documentation makes a specific promise about price: it passes through the providers' token rates with no markup on inference. We checked that promise against the providers' own price pages on September 28, 2026, and for the closed models it holds exactly — Claude Sonnet 5, Claude Opus 5.5, GPT-6 Luna, GPT-6 Sol and Gemini 3.8 Flash cost the same per token on OpenRouter as they do direct.

So the middleman tax is not in the token price. It is in three other places: a 5.5% fee when you buy credits (with a $0.80 minimum), an 8% fee on the Business plan, and a 5% fee on bring-your-own-key usage above $25,000 a month. And for open-weight models the story flips: third-party hosts on OpenRouter can undercut the model maker's own API by more than the fee.

This guide prices all three cases with the published rates, so you can see which one applies to your bill.

The rate cardsSame token prices for closed models

Cache prices matched as well where we could compare them like for like: $0.20 per million cached reads and $2.50 per million cache writes on Sonnet 5, and $0.01 cached input on GPT-6 Luna. The rate-card mechanics themselves — cache writes, batch discounts, long-context bands — are unchanged by the router, and are decoded in our Claude API pricing guide and OpenAI API pricing guide.

The fee5.5% on credits, and the $0.80 floor

On the Standard pay-as-you-go plan, OpenRouter charges 5.5% when you buy credits by card, with a minimum of $0.80, or 5% when you pay in crypto. The fee is taken at purchase, not per request — which makes the effective rate depend on how much you top up at once.

Two terms deserve a note before you prepay a large balance. OpenRouter's FAQ says refunds of unused credits can be requested only within 24 hours of the purchase, and that it reserves the right to expire unused credits one year after purchase. Buy in amounts you will actually spend within the year.

The Business plan, which adds EU/US in-region routing and workload identity federation, charges an 8% platform fee instead of 5.5%. Enterprise contracts list "fee discounts available" without a number.

What it costs per monthCredits vs bring-your-own-key

The alternative to buying credits is BYOK: you attach your own OpenAI, Anthropic or Google key, the provider bills you directly at its normal price, and OpenRouter charges a fee of 5% of what that usage would have cost on OpenRouter — but only above $25,000 of BYOK usage a month on pay-as-you-go ($200,000 on Enterprise).

At hobby scale the fee is noise: $2.75 on a $50 month. At $5,000 a month on credits it is $3,300 a year — real money for a small team, and entirely avoidable with BYOK, which is free at that volume. Above $25,000 BYOK still roughly halves the fee: $3,750 instead of $5,500 on a $100,000 month.

Open-weight modelsWhere the router can be cheaper

For closed models there is one price and OpenRouter can only add to it. Open-weight models are different: anyone can host them, so OpenRouter lists many providers for the same model at different prices. DeepSeek's V4.1 Flash is the clearest current case.

DeepSeek's own API charges $0.30 per million input tokens and $1.20 per million output tokens at peak hours and half that off-peak. On OpenRouter on September 28, 27 endpoints served the same model, from $0.025 to $0.375 per million input tokens. DeepSeek's own endpoint there listed $0.30/$1.20 — its peak rate.

Even with the 5.5% fee, the cheapest fp8 host costs 44% less than DeepSeek off-peak and 72% less than DeepSeek at peak on this workload. That is the opposite of a middleman tax — but it comes with conditions that the price column does not show.

For a deeper look at DeepSeek's own pricing windows and cache behaviour, see our DeepSeek vs OpenAI API cost comparison.

What the fee buysReasons to pay it anyway

A fee is only a tax if you get nothing for it. What OpenRouter adds that a direct key does not:

• One key, one bill. Switching a request from Claude to GPT to Gemini is a model-name change, not a new account, contract and invoice.
• Fallbacks across clouds. Sonnet 5 is served through Anthropic, AWS, Google and Azure at the same list price; if one endpoint is down, routing moves to the next.
• Price competition on open models. As the DeepSeek table shows, the marketplace can more than pay for the fee on open-weight workloads.
• Free models for experiments. 50 requests a day with no credits, 1,000 a day after buying $10 — useful for prototypes, not for production.

What it does not add: lower prices on closed models, provider-specific features before OpenRouter supports them, or a contract with the model maker itself.

The decisionWhen to use which route

Before choosing, price your actual monthly volume. The AI Cost Calculator compares models at list price for your input, cached input and output mix; add 5.5% for credits or 0% for BYOK under $25,000 to get the router's figure. For agents whose context grows with every step, use the AI Agent Cost Calculator — a per-call estimate understates both the bill and the fee.

Frequently asked questions

Does OpenRouter mark up API prices?

Not the token prices. OpenRouter states it passes through provider pricing with no markup on inference, and on September 28, 2026 the per-token rates for Claude, GPT-6 and Gemini models matched the providers' own pages. It charges instead a 5.5% fee when you buy credits by card (5% in crypto), with a $0.80 minimum.

Is OpenRouter cheaper than using the OpenAI or Anthropic API directly?

For closed models, no: the tokens cost the same and credits add 5.5%. With your own key (BYOK) it costs the same as direct up to $25,000 of usage a month. For open-weight models it can be cheaper, because third-party hosts compete below the model maker's price.

What is OpenRouter's BYOK fee?

On pay-as-you-go plans the first $25,000 of BYOK usage each month carries no fee; beyond that OpenRouter charges 5% of what the same usage would cost on OpenRouter. Enterprise plans include $200,000 a month before the fee applies. The provider bills you directly for the tokens.

Do OpenRouter credits expire?

OpenRouter's FAQ says it reserves the right to expire unused credits one year after purchase, and that refunds of unused credits can be requested only within 24 hours of the transaction.

Why is the same model priced differently on OpenRouter?

Different providers host it. Closed models are priced by their maker, with occasional regional endpoints about 10% higher. Open-weight models are hosted by many companies, each setting its own price and often running different precision (fp8, fp4) and output limits.

Sources and methodology

OpenRouter's fees, plans, BYOK allowances, free-model limits, credit expiry and refund terms were read on September 28, 2026 from openrouter.ai/pricing and the OpenRouter FAQ; routing parameters from its provider selection guide. Per-model and per-host prices, precision and output limits were read the same day from OpenRouter's public model and endpoint API. Direct prices are the providers' own: Anthropic and OpenAI as documented in our pricing guides and re-compared against OpenRouter's listing, Google from its Gemini API pricing page, and DeepSeek from its Models & Pricing page.

Every fee, monthly total and host comparison is reproduced in drafts/openrouter-vs-direct-api-pricing-numbers.mjs. The DeepSeek workload is an illustrative shape, not a measured benchmark, and prices only tokens — it assumes every host produces acceptable output, which you should verify. Host prices on a marketplace change often; re-check the endpoint list before committing volume.