OpenRouter Pricing 2026: Credit Fees, BYOK, Free Model Limits, and What Routing Costs Next to Going Direct

Published 29 October 2026. Every OpenRouter fee, limit and price below was read from openrouter.ai, its documentation or its public models API on 29 October 2026, and is as of that date.

OpenRouter charges for models at the provider's own list price and makes its money on the credits you buy: a 5.5% fee on credit purchases on the self-serve Standard plan (with a $0.80 minimum on card payments, and 5% for crypto), 8% on the Business plan, and a negotiated fee on Enterprise, according to its pricing page and FAQ. OpenRouter's own wording is that inference "is billed at the provider's list price on every plan" and that the fee is "charged when you buy credits... never on individual requests". If you bring your own provider keys (BYOK), the first $25,000 of list-price inference each month carries no OpenRouter fee on Standard and Business, and usage above that costs 5% of what the same request would have cost on OpenRouter. Free models are capped at 20 requests a minute and 50 a day, rising to 1,000 a day once an account has bought at least $10 of credits.

Most of the pages that rank for "openrouter pricing" describe a structure that has since changed. They say BYOK is free for the first million requests a month; the pricing page and BYOK docs now measure the allowance in dollars of list-price inference, "not request count". They describe three tiers, when there are now four, with a self-serve Business plan at 8%. And some say prompt caching and batch discounts are unavailable through OpenRouter, when its docs now document provider-sticky caching and a Batch API at "typically 50%" of the synchronous price.

In the worked example further down, a 40-turn conversation on Claude Sonnet 5.5 with a warm cache costs $0.240 direct from Anthropic, $0.253 through OpenRouter on Standard, $0.259 on Business and $0.240 plus no fee through BYOK, which is the whole argument in four numbers: the token price is the same, and what you pay for is the fee on the credits.

OpenRouter plans: Free, Standard, Business and Enterprise

From OpenRouter's pricing page, as of 29 October 2026.

Standard and Business bill the same way, and OpenRouter says "the only billing difference between the two is the fee on credit purchases". Both have no minimum spend and no lock-in, an upgrade to Business is self-serve from account settings with "no contract, no minimum spend, and no sales conversation", and you can downgrade 30 days after upgrading. Enterprise is priced "by contract and based on volume, prepaid credits and annual commitments, with invoicing".

Does OpenRouter mark up model prices?

No, and the claim can be checked line by line. OpenRouter's public endpoints API lists every provider that serves a model with that provider's price, and on 29 October 2026 the first-party endpoint for each model below listed exactly the price we read from the provider's own page for our direct-pricing guides on 27 and 28 October.

US dollars per million tokens, from https://openrouter.ai/api/v1/models/{model}/endpoints, read 29 October 2026.

Two details sit underneath "no markup". The first is regional endpoints. Several providers list a region-pinned endpoint next to their standard one, and on 29 October those endpoints were priced 10% higher: Claude Sonnet 5.5 on amazon-bedrock/us-east-1, google-vertex/us and azure/us at $2.20 input and $11.00 output, Mistral Small 4 on mistral/eu at $0.165 and $0.66. OpenRouter's in-region routing docs say "some providers bill a premium for serving a model from a specific region", and that default routing on the global domain skips those endpoints whenever a standard-priced one exists, so you pay the premium only when you pin a region or use the EU or US domains on Business.

The second is that "the provider's list price" means the price of whichever provider serves the request, and for an open-weight model that is often not the lab that trained it. The section on routing below shows what that does to a bill.

OpenRouter also says what happens when a price moves: it keeps routing to the model and "you will be charged at the new rate". A deprecated model returns a 404 with "no endpoints for this model found", so pin an explicit model ID if you need a stable version.

The credit purchase fee: 5.5%, an $0.80 minimum, and what is refundable

Standard and Business accounts buy credits in US dollars, manually or with auto top-up, and each request deducts its cost from the balance. The fee applies at purchase. OpenRouter's FAQ states the card fee as "5.5% ($0.80 minimum)" and the crypto fee as 5%, and the pricing page states 8% for Business.

The minimum matters only for small top-ups. Below $14.55 the $0.80 floor is more than 5.5%, so a $5 top-up pays a 16% fee and a $10 top-up pays 8%; from about $15 upward you pay the headline rate. A team spending $2,000 a month on Standard pays $110 a month in fees, and $160 on Business.

The rest of the billing rules, all from the FAQ and pricing page:

• Refunds. Unused credits can be refunded within 24 hours of the purchase through the Credits page; after that they are non-refundable. "The platform fees are non-refundable", and crypto payments "are never refundable".
• Expiry. OpenRouter's terms reserve the right "to expire unused credits after one year of purchase". Deleting an account forfeits its credits.
• Taxes. Prices exclude tax, and VAT or GST "may be added on invoices" where required.
• Payment methods. Cards, AliPay and USDC on Standard and Business; invoicing and purchase orders on Enterprise.
• Streaming. Billed the same as non-streaming, per token.
• A 1% discount. If you opt in to letting OpenRouter use your prompts and completions to improve the product, you get "a 1% discount on all model usage", according to its data collection page. It is off by default, and for most teams a 1% saving is not worth handing over production prompts.

New accounts receive what the FAQ calls "a small free allowance to test out OpenRouter"; OpenRouter does not publish the amount.

Standard at 5.5% or Business at 8%

Business costs 2.5 points more on every credit purchase, and token prices do not change. What the extra fee buys, per the pricing page, is EU and US in-region routing (requests to eu.openrouter.ai are "decrypted in the EU and routed only to shared provider endpoints in Europe"), Workload Identity Federation so services authenticate with tokens from your own identity provider instead of long-lived keys, advanced notifications, up to 1,000 workspaces instead of 5, and a signed copy of the DPA on request.

For a team spending $10,000 a month, Business costs $250 more a month than Standard. That is easy to justify when a customer contract requires EU processing, and hard to justify otherwise. Note also that the in-region domains route to regional endpoints, which carried the 10% premium shown above on the models we checked, so an EU-only workload on Business pays both the 8% fee and, on some models, a higher token price. The Sovereign AI page adds that "account, billing, and other operational records may be stored or processed in the US" even on the EU domain.

BYOK pricing: $25,000 a month free, then 5%

With bring-your-own-key, the provider bills you for inference directly at your own rates, and OpenRouter charges nothing until your BYOK usage passes $25,000 of list-price inference in a month on Standard or Business. Above that, the BYOK docs put the fee at "5% of what the same model/provider would cost normally on OpenRouter", deducted from your OpenRouter credits, so a BYOK account still needs a credit balance once it crosses the allowance. Enterprise allowances are custom.

The allowance is measured in dollars, not requests. A team running $40,000 a month of list-price inference through its own Anthropic and Google keys pays OpenRouter 5% of the $15,000 above the allowance, which is $750 a month, an effective 1.9%. The same $40,000 bought as Standard credits carries a $2,200 fee, and $3,200 on Business. A team under $25,000 a month on BYOK uses OpenRouter's routing, fallback and logging at no OpenRouter fee at all.

Three BYOK behaviours change what you pay:

• Fallback to shared capacity spends credits. By default, if all your keys hit a rate limit or fail, OpenRouter falls back to its own shared endpoints, billed to your OpenRouter credits at the normal rate plus the 5.5% or 8% you paid for them. Each key has a "Shared capacity fallback" setting that can turn this off for the models the key covers or for the whole provider.
• Budgets ignore BYOK by default. Guardrail budgets count only OpenRouter credit spend unless you enable "Include BYOK spend", so a budget can look far from its cap after heavy BYOK use.
• Regional premiums bill to you. With a prioritized BYOK key, a regional premium is billed to your own provider account rather than to OpenRouter credits.

BYOK is also the answer to the question incumbents raise about negotiated rates: if you have a committed-spend discount with a provider, BYOK keeps it, and OpenRouter's enterprise page markets exactly that ("bring your own provider keys and keep your negotiated discounts").

OpenRouter free models and their rate limits

From OpenRouter's limits documentation and FAQ, as of 29 October 2026.

A free model is any model ID ending in :free. The pricing page advertises "25+ free models" from "4 free providers"; OpenRouter's public models API listed 16 :free text models on 29 October, among them Google's Gemma 4 31B and Gemma 4 26B A4B, NVIDIA's Nemotron 3 Ultra and Nemotron 3.5 Super VL, Cohere's North Mini Code and Poolside's Laguna S 2.1. The Free Models Router (openrouter/free) picks one for you at no charge.

The tier is set by credits purchased over the account's lifetime, so a single $10 top-up lifts the daily cap twentyfold for good. Four rules catch people out. Failed attempts count toward the daily cap. A negative credit balance blocks free models until you add credits. The Free plan has no prompt caching, budgets or auto-routing, per the pricing table. And a fallback from a free model to a paid one is billed: the free variant docs warn that if a :free model sits in a models fallback array, "a fallback to a paid model incurs charges even though the first model was :free". Free models are also a poor fit for long conversations; at 50 requests a day, one 40-turn chat uses most of the day.

On paid models there is no plan-based rate limit. A 429 normally comes from the provider, and fallback routing retries other providers for the same model before the error reaches you; only "very high burst traffic" meets OpenRouter's Cloudflare DDoS protection. The limit that does bite on paid models is the in-flight spending budget: because a request is charged when it finishes, OpenRouter estimates each request's cost up front from the input and max_tokens, and rejects new requests with a 402 when the running total would exceed a fraction of your balance. Setting max_tokens sensibly and keeping a healthy balance avoids it.

How OpenRouter routing changes what you pay

When you leave the provider choice to OpenRouter, the provider routing docs say the default is to "load balance requests across providers, prioritizing price": among stable providers, the cheaper ones are weighted by the inverse square of their price, so a $1 provider is nine times as likely to be tried first as a $3 one. A separate page on how traffic is split describes the weighting as an "adjusted price" built from the provider's cache-adjusted effective price over the last 30 minutes, time to first token, reliability, 429 rate and speed, so a faster endpoint "can carry a higher adjusted price and stay competitive". Either way, default routing is price-aware, not cheapest-first.

Three switches turn it into a price you choose:

• provider.sort: "price" always takes the cheapest endpoint, with no load balancing.
• :floor on the model ID does the same and also admits a provider's discounted flex tier, billed at the flex rate when flex serves the request.
• :nitro sorts by throughput and admits priority tiers, and a request served on a priority tier is "charged that endpoint's priority pricing".

Naming providers, using BYOK or sorting all switch off the default load balancing.

For a closed model such as Claude or Gemini, every endpoint on our list charged the lab's price or the 10% regional price, so routing barely moves the bill. For an open-weight model it moves it a great deal, in both directions. On 29 October OpenRouter listed 42 endpoints for DeepSeek V4.1 Flash 1017. DeepSeek's own endpoint charges $0.15 input and $0.60 output off-peak and double at peak; several third-party hosts list a flat $0.30 and $1.20 around the clock, which is DeepSeek's peak rate at every hour, while others go as low as $0.0133 input on an endpoint labelled fp4. GPT-OSS 120B had 25 endpoints, from $0.03 input and $0.15 output on fp4 and int4 endpoints up to $0.35 and $0.75, against Groq's $0.15 and $0.60.

The cheapest endpoints are frequently quantized, and the endpoints API labels each one (fp4, mxfp4, int4, fp8, bf16), so a request routed purely on price may run a 4-bit version of the model you named. Whether that costs quality on your task is something only your own evals can settle, and the provider routing docs let you filter on quantization if it does. Our Kimi K3 pricing guide walks through the same pattern on K3's 30 endpoints.

Prompt caching, Batch and the costs that sit outside token prices

Prompt caching works through OpenRouter, at each provider's cache prices, on Standard and above. OpenRouter's prompt caching guide describes provider sticky routing: after a cached request, later requests in the same conversation go to the same provider endpoint, keyed on a hash of the opening messages or on a session_id you send. Sticky sessions "expire after 10 minutes of inactivity", and sticky routing switches on only when the provider's cache read is cheaper than its input price. Anthropic models still need cache_control breakpoints, and Claude cache writes cost more than plain input, which the response reports as a negative cache_discount. If you set provider.order manually, sticky routing is off and cache hits depend on your ordering.

Batch costs about half. The Batch API bills "typically 50% of the model's standard per-token price" and returns results within 24 hours on most providers (Google Vertex can take up to 96). It picks "the cheapest provider that serves the model in batch mode", keeps data for 30 days and therefore cannot run under Zero Data Retention, runs on US infrastructure, and does not support OpenRouter web search. The models API lists batch variants such as anthropic/claude-sonnet-5.5:batch at $1.00 input and $5.00 output.

Identical requests can be free. Response caching returns a stored response for an identical request "with no billing", which suits test suites and repeated lookups; two identical requests arriving at once are both billed.

Failed requests are not billed. Under Zero Completion Insurance, a response with no output tokens and a blank finish reason, or with an error finish reason, deducts nothing for prompt, completion or reasoning tokens, "even if the underlying provider charges for prompt processing". Routing through openrouter/auto adds no fee.

Web search bills on top. The web search plugin charges $0.007 per request for Exa search in its default auto mode, including up to 10 results, and $0.001 for each extra result; Firecrawl costs $0.005 per search plus $0.0125 per result; models with native search pass the provider's search fee through (the models API lists a web_search charge of $0.01 on Claude Sonnet 5.5 and $0.014 on Gemini 3.8 Flash). Search and file parsing charges can still apply when the model's response fails.

What a 40-turn conversation costs direct and through OpenRouter

Take the chat assistant from our other pricing guides: a 2,000-token system prompt, 150-token user messages and 350-token replies, so a 40-turn conversation sends 476,000 input tokens and returns 14,000 output tokens. The direct figures below are the ones published in our Claude, Gemini, DeepSeek, Groq, Mistral and Kimi guides, each with its own provider's caching rules. Since OpenRouter's first-party endpoint price matched every one of them on 29 October, the OpenRouter figures are the direct figure plus the credit fee.

Direct prices read from each provider on 27 or 28 October 2026 and matched against OpenRouter's endpoints API on 29 October; OpenRouter columns add the 5.5% or 8% credit fee. With BYOK under the $25,000 monthly allowance, the OpenRouter cost equals the direct cost.

Two assumptions hold the table together. The cache-warm columns assume OpenRouter keeps the conversation on one provider, which its sticky routing is built to do, and that the gap between turns stays under the shorter of OpenRouter's 10-minute sticky window and the provider's own cache lifetime. And the request lands on the first-party endpoint; if default routing sends DeepSeek traffic to a flat-rate host at $0.30 and $1.20, the off-peak conversation costs what the peak one does, and if it sends it to a 4-bit host it may cost less.

At 10,000 such conversations a month with a warm cache, Claude Sonnet 5.5 costs $2,396 direct, $2,528 through OpenRouter on Standard and $2,588 on Business, so the fee is $132 or $192 a month. On Gemini 3.8 Flash the same volume costs $1,110 direct and $1,171 on Standard, and on deepseek-flash off-peak it costs $130 and $137. In absolute terms the fee is small on cheap models and real on expensive ones, and in percentage terms it never changes.

When OpenRouter pays for itself

OpenRouter costs 5.5% of model spend on Standard, so it pays for itself when it saves more than that, in money or in engineering time. Going by the numbers above, it does in five situations.

1. BYOK under $25,000 a month. You pay your providers directly and OpenRouter charges nothing, so the gateway, the fallback and the unified logs are free. This is the cheapest way to use OpenRouter by a distance, and the one most pricing guides miss.
2. Open-weight models where a cheaper host is acceptable. If your evals pass on a third-party or quantized endpoint that costs 10% or more below the lab's own price, routing saves more than the fee. On GPT-OSS 120B the cheapest endpoints listed one-fifth of Groq's input price and a quarter of its output price on 29 October.
3. Several providers, one integration. Calling Anthropic, Google, DeepSeek and Moonshot through one OpenAI-compatible API, one key and one balance replaces four integrations, four billing relationships and four sets of rate-limit handling. At $2,000 a month of spend the fee is $110; one engineer-day of integration and on-call work costs more than that.
4. Fallback you would otherwise build. Provider failover with automatic retry, plus no charge for errored or empty responses under Zero Completion Insurance, is reliability work you do not write yourself.
5. Prototyping and model selection. Free models, the Free Models Router and per-request model switching make OpenRouter the cheap place to decide which model to buy.

It does not pay for itself when you run one provider at high volume and do not need fallback (the fee is pure overhead, and BYOK above $25,000 is the cheaper way to keep OpenRouter in the path), when you need batch under Zero Data Retention or regional batch, which OpenRouter's Batch API does not support, or when default routing sends a time-priced model like deepseek-flash to flat-rate hosts during DeepSeek's off-peak hours. In that last case, pin the provider with provider.only or provider.order and the problem goes away.

OpenRouter Enterprise pricing

Enterprise has no list price. The pricing page describes contract pricing "based on volume, prepaid credits and annual commitments, with invoicing", with "discounted OpenRouter fees" and a custom BYOK allowance, and states plainly: "We do not discount inference." What Enterprise adds over Business is SSO (SAML) and SCIM, managed policy enforcement, contractual SLAs, private endpoints as an add-on, custom DPA terms, invoicing and purchase orders, and a support team with "a support SLA with committed response times, a dedicated Slack channel", a customer success manager and a forward-deployed engineer. OpenRouter's enterprise page also lists SOC 2 Type II.

For a CIO the practical reading is that Enterprise negotiates the platform fee and the BYOK allowance, not token prices; any token discount comes from your own contracts with the providers, carried in through BYOK.

How to keep an OpenRouter bill down

A request through OpenRouter costs the serving endpoint's token price (cached input, uncached input, output and reasoning), plus any search or tool charges, and the credits that paid for it carried a 5.5% or 8% fee. Six levers move that total, roughly in order of size:

1. Use BYOK for your main providers and stay under, or knowingly above, the $25,000 monthly allowance.
2. Keep the prefix stable and send a session_id, so sticky routing keeps the cache warm, and add cache_control on Claude.
3. Send work nobody waits for through the Batch API at about half price.
4. Pin or sort providers deliberately on open-weight models, with provider.sort: "price" or :floor where your evals accept the cheaper endpoints, and provider.only where they do not.
5. Top up in amounts above $15, so the $0.80 minimum never applies, and set per-key credit caps and workspace budgets.
6. Stop resending the whole conversation once chats run long. Every row in the table above grows with the square of the turn count, and no fee or cache changes that shape.

We build Maximem Synap, a memory layer that returns a short, ranked set of statements about the user, trimmed to a token budget you set, in place of the full history, which keeps each request's input roughly flat as a conversation grows, whichever provider OpenRouter routes it to.

Frequently asked questions

How much does OpenRouter cost?

OpenRouter bills model tokens at the serving provider's list price and charges a fee when you buy credits: 5.5% on the Standard plan (with a $0.80 minimum on card payments, and 5% for crypto), 8% on the Business plan, and a negotiated fee on Enterprise, as of 29 October 2026. The Free plan has no fee and gives access to free models only.

Does OpenRouter mark up API prices?

No. On 29 October 2026 the first-party endpoints for Claude Sonnet 5.5, Claude Haiku 5.5, Gemini 3.8 Flash, DeepSeek V4.1 Flash, Mistral Small 4 and Large 4, Kimi K3 and K3.1 Pro, and GPT-OSS 120B on Groq all listed the same price on OpenRouter as on the provider's own pricing page. Region-pinned endpoints cost 10% more on the models we checked, and default routing skips them when a standard endpoint exists.

What is the OpenRouter BYOK fee?

On Standard and Business, the first $25,000 of list-price inference per month through your own provider keys carries no OpenRouter fee. Above that, OpenRouter charges 5% of what the same request would cost on OpenRouter, deducted from your credits. The allowance is measured in dollars, not request count, and Enterprise allowances are custom.

What are OpenRouter's free model rate limits?

Free models (IDs ending in :free) allow 20 requests per minute and 50 requests per day, rising to 1,000 requests per day once the account has bought at least $10 of credits in total. Failed attempts count toward the daily cap, and a negative balance blocks free models.

Does OpenRouter have rate limits on paid models?

OpenRouter sets no plan-based rate limit on paid models. Limits come from the providers, and fallback routing retries other providers for the same model automatically. OpenRouter's own DDoS protection rejects only very high burst traffic, and an in-flight spending budget can return a 402 when many expensive requests run at once.

What is the difference between OpenRouter Standard and Business?

Token prices are identical. Business charges 8% on credit purchases instead of 5.5%, and adds EU and US in-region routing, Workload Identity Federation, advanced notifications, up to 1,000 workspaces and a signed DPA on request. Upgrading is self-serve with no contract or minimum spend.

Does prompt caching work through OpenRouter?

Yes, on Standard, Business and Enterprise. Cached tokens bill at the serving provider's cache rates, OpenRouter keeps a conversation on the same provider with sticky routing that expires after 10 minutes of inactivity, and Claude models still need cache_control breakpoints.

Does OpenRouter have a Batch API discount?

Yes. Batch requests are typically billed at 50% of the model's synchronous price, with results within 24 hours on most providers. Batch does not support Zero Data Retention, regional routing or OpenRouter web search.

Are failed requests billed on OpenRouter?

No. A response that errors, or returns no output tokens with no finish reason, is not charged for model tokens under Zero Completion Insurance, so with fallback you pay only for the attempt that answers. Web search or file parsing that already ran may still be billed.

Is OpenRouter cheaper than going direct?

For closed models it costs the fee more than going direct, unless you use BYOK under $25,000 a month, where it costs the same. For open-weight models it can cost less, because third-party and quantized endpoints often undercut the lab's own price by more than the 5.5% fee. A 40-turn Claude Sonnet 5.5 conversation with a warm cache costs $0.240 direct and $0.253 through OpenRouter on Standard.

Do OpenRouter credits expire, and can I get a refund?

OpenRouter reserves the right to expire unused credits one year after purchase. Unused credits can be refunded within 24 hours of the purchase; the platform fee is non-refundable and crypto payments are never refundable.

Sources

All retrieved 29 October 2026.

• OpenRouter: pricing, enterprise, FAQ, limits, BYOK, provider routing, traffic routing, free variant, floor variant, nitro variant, Free Models Router, prompt caching, Batch API, response caching, Zero Completion Insurance, web search, in-region routing, Sovereign AI, data collection, and the public models API with its per-model endpoints listings
• Direct prices: our guides to Claude, Gemini, DeepSeek and Groq API pricing (read from each provider on 27 October 2026) and Mistral and Kimi K3 API pricing (28 October 2026), each citing the provider's own pricing page

Worked examples are our own arithmetic on those rates.