OpenRouter pricing has two layers. You pay each model's token rate, passed through at the provider's published price with no markup, and OpenRouter adds a platform fee when you load credits. Its published fee schedule sets that fee at 5.5%, with a $0.80 minimum per purchase and 5% for crypto.

Bring your own provider key, and the fee falls to 5% of the normal cost after the first million requests a month. Free model variants cost nothing but throttle you to about 20 requests a minute. The rest of this page shows the math, the catches, and how to keep the spend under control as a FinOps for AI problem rather than a billing surprise at the end of the month.

How OpenRouter Pricing Works?

OpenRouter prices every request along two independent lines, and the bill at the end of the month is just the sum of the two:

• Token cost, charged per million input and output tokens at the rate the underlying model provider sets, with nothing added on top.

• Platform fee, a 5.5% skim on the money you load as credits (5% for crypto), with a $0.80 minimum per top-up.

That split is why one credit balance can pay for 300+ models from OpenAI, Anthropic, Google, Mistral, DeepSeek, and others without OpenRouter publishing a price list of its own. In practice it behaves like a multi-provider LLM cost management tool with a checkout attached, where the model picks the token price and OpenRouter picks the fee.

Confuse the two layers and the forecast misses every month. The token line is the bigger one for most teams, which is why the model you route to moves the bill far more than the platform fee ever will, and the unit math behind that decision is the heart of token economics.

How OpenRouter Makes Money

OpenRouter runs a marketplace take-rate. It does not mark up the tokens. As its stated operating principle puts it, the platform passes provider rates through untouched and instead skims a percentage of the money you load as credits. The more you spend on inference, the more fee revenue it earns, which is why the headline rate stays flat while the dollar figure grows with usage.

So what you actually pay breaks down into three buckets:

• Token cost, set by the model provider, the largest line for most teams.

• Credit fee, the 5.5% (or 5% crypto) skim when you top up.

• BYOK overage, a smaller 5% fee that only applies past the free request tier.

For a buyer this is mostly good news. The rate does not change because you picked an expensive model. The one trap is the minimum fee, which makes tiny top-ups costly in percentage terms.

The 5.5% Credit Fee Explained

When you add credits with a card, OpenRouter charges 5.5% plus a $0.80 floor on the transaction. The floor is what stings small users. A $5 test top-up still pays the $0.80 minimum, which works out to 16%, while a $100 top-up pays $5.50, the true 5.5%. Crypto payments are charged 5% instead.

Here is the effective rate by top-up size, which no model catalog shows you in one place:

A worked example makes the rule obvious. Top up $5 four times in a month, and you pay $3.20 in fees on $20 of credit, a 16% drag. Load that same $20 in one purchase, and you pay $1.10, a 5.5% drag. The practical rule: add credits in larger, less frequent amounts so the floor never dominates.

Model Token Pricing: Pass-Through, No Markup

OpenRouter's per-model rate card lists each model at the provider's own rate, with no markup, so a call costs the same per token as going direct. The myth that it inflates Claude or GPT prices is wrong for current models, since the catalog rate matches the number the provider publishes.

Because tokens drive most of the bill, routing decisions dwarf the platform fee by a wide margin. A frontier model can cost ten to thirty times a small model for the same task, so handing simple work to a cheaper model saves far more across a month than the 5.5% ever takes from you. The GPT family on the platform bills at the same numbers documented for OpenAI API pricing.

Claude works the same way through the gateway. OpenRouter forwards input and output tokens completely unchanged, so there is no penalty for reaching the model through OpenRouter instead of going to the provider directly. The smart move is to pick the model that fits the job rather than the logo on it, and the separate input and output rates stay identical to those under Anthropic API pricing.

Google's models follow the same rule with no added margin layered on top. Across all three major vendors, the pattern is identical, where the provider sets the token price, OpenRouter passes it straight through, and your only platform cost is the credit fee you already paid at top-up. The published rates that flow through are the ones listed for Gemini API pricing.

The catalog runs well beyond the big three. From one key, you can route to a wide mix of models:

• xAI Grok for general chat and reasoning.

• DeepSeek R1 for low-cost reasoning, with cache-hit and off-peak discounts.

• Mistral models for cheaper European-hosted inference.

• 28+ free variants for prototyping at zero token cost.

Grok is a useful reminder to read the rate before you route, since its tiers shift with model size, and a heavier variant costs noticeably more for the same prompt. Matching the model tier to the task, rather than defaulting to the largest option, is the single biggest saving on the whole platform, and the tiers are spelled out under Grok API pricing.

DeepSeek shows the same idea from the discount side of the ledger. Its cache-hit and off-peak rates reward batching and prompt reuse, so a team that schedules its heavy jobs into the cheaper overnight window can cut a reasonable bill by a wide margin without switching models at all. Those exact tactics, from cache reuse to off-peak scheduling, are what the DeepSeek cost optimization tools roundup walks through for production teams.

BYOK Pricing: Bring Your Own Key

BYOK lets you attach your own provider keys so inference bills your provider account directly, while OpenRouter still handles routing. The first million BYOK requests each month are free. Past that, the BYOK documentation sets a fee of 5% of what the same model and provider would normally cost through credits, deducted from your balance.

BYOK suits teams that already hold negotiated provider rates or committed-use discounts. Below a million requests a month, it removes the platform fee entirely. Above that, you pay 5% on the overage, so model the volume before you assume BYOK is free.

Free Models and Rate Limits

OpenRouter offers free model variants, marked with a :free suffix, that cost $0 for input and output tokens. The catch is throughput. Free variants are throttled to about 20 requests per minute, and the daily cap depends on lifetime credit purchases.

Its rate-limit rules split the daily cap in two:

• Under $10 in lifetime credits: 50 free requests per day.

• $10 or more, ever: 1,000 free requests per day, permanently.

The 429 errors people hit on free models almost always come from this daily limit, not a bug in your code. For paid search-grade work, lower-cost retrieval models often undercut a frontier model on the very same task, including those listed under Perplexity API pricing.

Pay-as-You-Go vs Enterprise

The default plan is pay-as-you-go with no minimum and no lock-in. You load credits, optionally set auto top-up, and pay only for what you use. For most teams, this is the right starting point and needs no sales call to begin.

The Enterprise plan adds invoicing, purchase orders, and discounted platform fees, as its enterprise setup guide lays out. It does not discount the inference, since those rates belong to the providers. Enterprise makes sense at high, steady volume where procurement and a lower fee matter more than flexibility.

OpenRouter vs Direct API: When the Fee Is Worth It

The decision comes down to spend and how many providers you run. Under a few hundred dollars a month across several models, the 5.5% buys one key, one balance, and instant model switching, worth more than the few dollars it costs.

At high production volume on a single model, the fee turns into real money and direct keys usually win on both cost and rate limits. A side-by-side LLM cost comparison across providers usually moves the bill more than the gateway fee ever does.

A quick decision rule:

Run the actual numbers before committing, since the break-even shifts with your model mix. Teams standardizing on Google, for instance, often find that the direct route paired with the levers in Vertex AI cost optimization tools beats the gateway once volume climbs into production territory.

How to Forecast and Control OpenRouter Spend

A single credit balance across 300+ models is convenient and hard to attribute. You cannot tell which team, feature, or model burned the credits without instrumentation. This is where most OpenRouter bills go sideways, since spend shows up only as a shrinking balance, not as a cost per workload you can defend in a review.

Treat it like any other variable cloud cost from day one. Start by tagging every request and watching token volume per model, so you can see consumption building before the balance runs low. That request-level view, rather than a monthly invoice landing after the fact, is exactly what AI cost tracking tools are built to give a platform team.

Raw consumption data is only useful once someone can read it at a glance, and a wall of unlabeled token counts helps nobody on the finance side of the table. A flat, anonymous credit drain turns into a clear cost per team, per feature, and per model the moment you run it through AI cost visibility tools, which is the step most teams skip until the balance empties faster than the forecast promised.

Attribution is what makes the number defensible in a budget review. Map every request back to the product, team, or customer that triggered it, and the gateway bill stops being one mysterious figure that nobody can explain. That mapping is the job of dedicated LLM cost allocation tools that tie spend to the revenue it supports.

Guardrails come last but matter most of all. Set hard limits and budget alerts so a runaway agent or a bad deploy cannot quietly drain the balance overnight while everyone is asleep. A single alert on the credit balance is the cheapest insurance you can buy, and broader controls live in AI cost governance tools.

Conclusion

OpenRouter pricing is honest and simple once you separate the two layers. Tokens cost what the provider charges, and OpenRouter adds 5.5% on credit purchases with a $0.80 floor, dropping to 5% under BYOK above the free tier. The fee is a fair price for one key across every model at moderate scale. The real budget risk is not the 5.5%. It is losing sight of which workload spends the credits. Bring that spend into a FinOps practice and the gateway becomes a cost you control instead of a balance you keep refilling.

FAQs

Does OpenRouter markup model prices? 

No. OpenRouter passes through provider token rates with no markup, so a model costs the same per token as calling the provider directly. Its revenue comes from the credit-purchase fee, not the tokens.

What is the OpenRouter fee? 

OpenRouter charges 5.5% when you buy credits with a card, with a $0.80 minimum per transaction. Crypto payments are charged 5%. There is no per-token markup on top of that.

Why did a small top-up cost more than 5.5%? 

The $0.80 minimum fee. On a $5 top-up, the floor works out to an effective 16%. Add larger amounts less often so the fee stays near the 5.5% headline rate.

Is OpenRouter BYOK free? 

The first million BYOK requests each month are free. Above that, OpenRouter charges 5% of what the same model and provider would normally cost through its credits.

How many free requests does OpenRouter allow? 

Free models allow about 20 requests per minute. You get 50 requests per day under $10 in lifetime credits, rising permanently to 1,000 per day once you have purchased at least $10.

Is OpenRouter cheaper than the direct API? 

For low to moderate multi-model spend, yes, because the 5.5% buys convenience and model variety. For high steady volume on one model, direct keys are usually cheaper and offer higher rate limits.

Is OpenRouter cheaper than OpenAI? 

Per token, no. OpenRouter passes GPT models through at OpenAI's published rates with no markup. Factoring in the 5.5% credit fee, OpenRouter ends up slightly more expensive than calling OpenAI direct on the same model.

Is OpenRouter cheaper than Anthropic? 

Same story. Claude models route through at Anthropic's published input and output token rates with no markup, and the 5.5% credit fee makes the total fractionally higher than calling Anthropic direct on the same model.

When should I use OpenRouter vs direct APIs? 

Use OpenRouter when you need many models from one key at low or mid spend, or for fast experimentation. Go direct when one model dominates your traffic at high volume, when rate limits matter, or when you have negotiated provider discounts.