OpenRouter Pricing (2026): When the 5.5% Fee Pays Back

OpenRouter is not cheaper than buying the same model direct: shared credits carry a 5.5% purchase fee, with a $0.80 minimum. The fee pays back only when one API, one bill, and provider fallback save more than $55 of operational work per $1,000 of model usage; BYOK changes the equation by carrying no OpenRouter fee through $25,000 a month.

OpenRouter's live pricing page was verified on August 13, 2026. The important choice is not Free versus paid. It is whether you fund one shared OpenRouter balance, bring provider keys you already own, or remove the gateway from the bill entirely.

OpenRouter pricing at a glance

OpenRouter has three routes, but only Pay-as-you-go has a public platform percentage. Free limits requests, Pay-as-you-go charges by model plus a funding fee, and Enterprise trades a public sticker for negotiated controls.

There is no public monthly or annual self-serve subscription. Pay-as-you-go is prepaid usage with no minimum usage commitment. That wording matters because the terms separately set a $5 minimum and $25,000 maximum credit purchase per transaction. Enterprise can involve volume commitments, but only a sales quote can reveal that contract's effective annual cost.

The Free plan is listed as an ongoing $0 route, not a timed trial. Pay-as-you-go is the default for a builder who wants paid models without opening accounts at every provider. Enterprise is justified by governance and support, not by a publicly provable token discount.

How OpenRouter billing works

OpenRouter bills the underlying model first and the account-funding method second. It says it passes through provider inference prices without a token markup, then charges 5.5% when credits are purchased. Both the chat interface and API draw down the same US-dollar credit balance.

Model pricing is not one number. Input tokens, output tokens, reasoning tokens, images, and per-request tools can each carry separate meters. One million tokens is the standard quote unit, so a model listed at $2 input and $6 output costs $0.002 per 1,000 input tokens and $0.006 per 1,000 output tokens.

The current catalog shows how wide the spread is:

Grok 4.6 and Qwen3.8 Max are both listed at $2 per million input tokens and $6 per million output tokens. Grok 4.6 was released on August 12, while Qwen3.8 Max was released on August 3. Their equal unit price removes budget from the first comparison; acceptance rate, latency, and reliability decide which one earns production traffic.

Claude Sonnet 5 is $2 input and $10 output per million. DeepSeek V4 Pro 0813 is $0.435 input and $0.87 output. GPT-5.6 Luna is currently marked 50% off at $0.10 input and $0.60 output, but the same model page lists higher prices on other provider endpoints. A promotional endpoint is a current price, not a durable budget promise.

That variability is the value and the risk of a router. OpenRouter can move a request among eligible providers for the same model, but the provider, caching, and discount path can change the effective charge. Pin the provider when price certainty matters more than fallback breadth, and verify the model page before committing a customer quote.

The 5.5% fee pays back in minutes, or it does not pay back at all

The $0.80 minimum makes small top-ups materially more expensive than the headline percentage. The pure 5.5% rate begins at about $14.55, because $0.80 divided by 5.5% is $14.5454. A $5 credit purchase carries an effective 16% fee. A $10 purchase carries 8%. At $15, the calculated fee is $0.825, which finally matches 5.5%.

That makes the smallest paid test a poor place to judge OpenRouter's steady-state economics. Stay on free models if they answer the test, or fund at least $15 when a paid model is required. Buying exactly $10 does raise the free-model daily cap from 50 to 1,000, but the minimum fee still dominates that transaction.

A normalized workload makes model cost easier to see. Assume 1,000 calls, each averaging 1,000 input tokens and 250 output tokens. That is 1 million input tokens and 250,000 output tokens in total.

At current listed rates, those 1,000 calls cost $3.50 on Grok 4.6 or Qwen3.8 Max, $4.50 on Claude Sonnet 5, $0.6525 on DeepSeek V4 Pro 0813, and $0.25 on the discounted GPT-5.6 Luna endpoint. Above the minimum-fee floor, adding 5.5% makes the shared-credit totals $3.6925, $4.7475, about $0.6884, and $0.26375 respectively.

The decisive metric is cost per accepted result, not cost per token. Divide token spend, funding fees, retries, and review time by the number of outputs that clear the workflow's quality bar. A model that costs half as much but doubles correction work is not cheaper.

The operational break-even is unusually small at low spend. At an explicit internal engineering-cost assumption of $100 per hour, the $5.50 fee on $100 of model usage equals 3.3 minutes. The $55 fee on $1,000 equals 33 minutes. The $550 fee on $10,000 equals 5.5 hours. If one bill, one API integration, and fallback handling save more time than that each month, OpenRouter pays back. If one provider serves a stable high-volume workload, direct billing wins.

The free plan is enough for evaluation, not a production promise

OpenRouter Free is useful when 50 requests a day covers the whole job. It includes 25+ free models across 4 free providers, costs $0 per token on those variants, and allows 20 requests per minute. A new user evaluating prompts, comparing a few open models, or running a personal utility may never need to pay.

The daily cap changes after a purchase. Accounts that have bought less than $10 of credits get 50 free-model requests per day. Once all-time purchased credits reach at least $10, the cap becomes 1,000 per day, while the 20-per-minute limit remains. This is a purchased limit increase, not a paid subscription.

Who never needs to pay? A solo builder whose non-critical evaluations fit below 50 daily requests, tolerate the available free-model catalog, and do not require a production support commitment can stay free indefinitely. The honest stopping point is any workflow that needs predictable paid-model access, more capacity, or accountability when a provider is constrained.

OpenRouter itself says free variants have low rate limits and are usually not suitable for production use. A customer-facing agent, revenue workflow, or scheduled batch should not depend on a free endpoint merely because yesterday's traffic fit under the cap.

BYOK changes the pricing decision

BYOK, or bring your own key, is OpenRouter's strongest price route for a company that already has provider accounts. Pay-as-you-go includes $25,000 per month of list-price inference with no OpenRouter BYOK fee. Enterprise includes $200,000 per month. Usage above either allowance carries a 5% fee.

The allowance is measured in list-price inference dollars, not request count. A Pay-as-you-go workspace with $50,000 of monthly BYOK inference has a $25,000 overage. Five percent of that overage is $1,250. The same account below $25,000 has no OpenRouter BYOK fee, although it still pays its providers directly.

BYOK changes more than the invoice. Prioritized provider keys are tried before OpenRouter's shared endpoints. If those keys fail or rate-limit, OpenRouter can fall back to shared capacity by default, which then consumes OpenRouter credits. Turning on Always use for this provider prevents that shared fallback when cost attribution or a provider contract matters more than resilience.

The budget control has a hidden switch. BYOK inference does not count toward guardrail or workspace budgets by default. A dashboard cap can therefore look comfortably below target while direct-provider spend accumulates outside it. Enable Include BYOK spend separately on guardrail and workspace budgets before finance relies on those totals.

For an existing multi-provider buyer under $25,000 a month, BYOK is the strongest OpenRouter deal: unified routing with no OpenRouter gateway fee. For a buyer who wants one prepaid balance and no provider accounts, shared credits earn the 5.5% premium through convenience. Above the BYOK allowance, compare the 5% overage with a zero-fee gateway or direct integration before the next billing cycle.

Hidden costs sit outside the token line

Credits can expire. OpenRouter's terms reserve the right to expire unused purchased credits 365 days after purchase. That makes a large speculative top-up an annual-lock risk even though there is no annual subscription. Fund observed usage, not an optimistic twelve-month forecast.

Refunds are narrower than most buyers expect. Unused credits can be requested back within 24 hours of the transaction. Platform fees are not refunded, and cryptocurrency payments are never refundable. Auto Recharge can prevent a production interruption, but it also turns an unnoticed traffic spike into repeated purchases, so pair it with per-key credit limits.

Web search is a separate meter on top of model tokens. OpenRouter currently lists Exa at $0.007, $0.012, or $0.015 per request depending on mode; Parallel at $0.001 or $0.005; and Perplexity search at $0.005. Firecrawl BYOK consumes Firecrawl credits, while native provider search passes through that provider's price. An agent that searches several times per answer can make tool calls more expensive than its text tokens.

The optional 1% usage discount also has a non-financial price. OpenRouter offers it when a user opts into prompt and completion logging. Its terms grant broad rights over opted-in logged content for service and commercial or business purposes, including anonymized licensing or sale. A 1% discount should not decide whether customer, source-code, or confidential prompts are logged.

Model pages can carry temporary discounts and provider-specific rates. GPT-5.6 Luna's $0.10/$0.60 endpoint is marked 50% off, while other listed endpoints are higher. Use current routing controls for experimentation, but store the verified model price beside every production budget and treat repricing as a trigger to review the route.

There is no public student discount, monthly seat, or annual self-serve discount on the OpenRouter pages verified in August 2026. The Free plan is the student and hobbyist route. Pay-as-you-go bills usage, and Enterprise pricing stays behind a quote.

OpenRouter versus Vercel AI Gateway, Requesty, and direct APIs

OpenRouter is not the cheapest gateway on the public platform-fee line. On $1,000 of list-price model usage, its 5.5% shared-credit fee is $55. Requesty applies a 5% model markup, or $50 on the same base usage. Vercel AI Gateway advertises $0 token markup and $0 platform fee, although payment-processing fees and optional feature charges can still apply. Direct provider billing also has no separate gateway fee.

OpenRouter earns its $55 through breadth and switching: 400+ models, 70+ providers, one balance, and configurable fallback. It is the strongest choice when a product genuinely moves among model families, when provider outages have business cost, or when a small team values one integration more than the percentage line.

Vercel AI Gateway wins the public pure-gateway price. Its Free tier includes $5 of monthly credit on eligible models, and its Paid tier sells provider tokens at list price without a platform fee. BYOK also carries no Vercel gateway fee, but it requires the Paid tier and purchased credits. The tradeoff moves into add-ons: Custom Reporting, team-wide provider allowlists, team-wide zero-data-retention enforcement, and trace drains have separate meters.

Requesty sits between them at a 5% markup with no subscription, seat fee, or minimum spend. Its Free route allows 200 requests a day on free models, while routing, caching, and EU data residency are included across plans. Choose it when those included controls fit the workflow and the $50 per $1,000 markup is easier to budget than a credit-purchase fee.

Direct provider access wins when one model owns most of the stable load. The gateway percentage then buys little, and an existing direct commitment may matter more. It loses when the team must recreate routing, budgets, logs, and failure handling across several vendors.

The decision rule is blunt: calculate the monthly gateway line, then make routing and billing automation earn that amount. If OpenRouter saves less than 5.5% of the workload's combined engineering and failure cost, choose Vercel, Requesty, or direct billing. If BYOK stays below $25,000, rerun the comparison with OpenRouter's gateway fee set to $0.

For a wider model-rate shortlist, use the cheapest AI API comparison. For one low-cost provider's current unit economics, the DeepSeek pricing breakdown separates cheap tokens from accepted outcomes.

The Monday move: price one accepted workflow

On Monday, route 100 sanitized requests from one bounded workflow for one week. Grok 4.6 and Qwen3.8 Max make a clean fresh comparison because both currently cost $2 input and $6 output per million tokens. The price line is held constant, so the test can expose the operational consequence: which route produces more accepted results with less retry and review work.

1. 1

   Define acceptance before routing

   Choose a result a reviewer can score consistently, such as a valid schema, passing code change, approved research brief, or correct classification. Remove customer, personal, and confidential data from the evaluation set.

2. 2

   Cap the trial

   Use one OpenRouter API key with a credit limit. If paid models are required, a $15 credit purchase reaches the pure 5.5% fee instead of the $0.80 minimum dominating a smaller top-up.

3. 3

   Hold price constant

   Send 50 requests to Grok 4.6 and 50 to Qwen3.8 Max with the same input and output caps. Record model cost, retries, latency, reviewer time, and accepted results in the same sheet.

4. 4

   Make the budget decision

   Review OpenRouter Activity by model, provider, and API key. Keep shared credits only if the routing and billing work saved exceeds the measured 5.5% fee. Move to BYOK, Vercel, Requesty, or direct billing when the gateway line buys less than it costs.

The Monday output is one number finance and engineering can share: cost per accepted result. It turns a model catalog into a budget decision and prevents a 5.5% convenience fee from becoming an unexamined permanent line item.

Frequently asked questions

How much does OpenRouter cost?

OpenRouter charges the selected model's listed input, output, and any tool rates. Shared credits add a 5.5% purchase fee with a $0.80 minimum; crypto funding carries a 5% fee. BYOK has no OpenRouter fee through $25,000 of monthly list-price inference on Pay-as-you-go or $200,000 on Enterprise, then charges 5% above the allowance.

Is OpenRouter free?

Yes. The Free route includes 25+ free models across 4 free providers. It is enough for light evaluation and personal utilities, but OpenRouter says the low-limit free variants are usually not suitable for production.

What is the daily limit for free OpenRouter models?

The limit is 50 free-model requests per day when all-time purchased credits are below $10, or 1,000 per day after at least $10 has been purchased. Both routes also have a 20-requests-per-minute cap.

How much do 1,000 tokens cost on OpenRouter?

Divide the model's per-million rate by 1,000. At $2 input and $6 output per million, 1,000 input tokens cost $0.002 and 1,000 output tokens cost $0.006, before allocating the separate credit-purchase fee.

Is OpenRouter cheaper than OpenAI?

Not for the same provider endpoint through shared credits. OpenRouter says it passes through provider token rates, then adds its credit-purchase fee. It can still be cheaper at the workflow level if unified billing and fallback save more labor or downtime than the fee; BYOK inside the free allowance removes the OpenRouter gateway fee.

Does OpenRouter offer a student discount?

No public student-specific plan or discount appears on the pricing, FAQ, or terms pages verified in August 2026. Students whose work fits 25+ free models and the daily request cap can remain on Free.

What is OpenRouter's refund policy?

Unused-credit refunds may be requested within 24 hours of purchase. Platform fees are non-refundable, and cryptocurrency payments are never refundable. Purchased credits may also be expired 365 days after purchase.

Did OpenRouter BYOK pricing change?

The live August 2026 policy is spend-based, not request-count based. Pay-as-you-go includes $25,000 of monthly list-price BYOK inference and Enterprise includes $200,000, with a 5% fee on usage above each allowance. Any request-count description is stale for the current plan.

Does OpenRouter have annual pricing?

No public annual self-serve plan is listed. Pay-as-you-go uses prepaid credits without a monthly or annual commitment, while Enterprise uses custom volume commitments and invoicing. The practical annual risk is unused credits that may expire after 365 days.

Get the AI Tools Map for Business Owners

Turn model prices into a practical adoption stack with a quality floor, routing role, and cost trigger for every tool. Subscribe to get the AI Tools Map free.