OpenRouter pricing: credit fees, BYOK and what you actually pay

Published on November 7, 2026 by Steven Jones

OpenRouter pricing has three separate parts: the AI model’s token charges, a fee when you purchase OpenRouter credits, and potentially a separate provider invoice if you bring your own API key (BYOK). Standard charges 5.5% on credit purchases, while Business charges 8%. Neither percentage is added to every model request.

For BYOK, Standard and Business currently include $25,000 of list-price inference per month without an OpenRouter usage fee. Charges above that threshold are 5% of the excess. This guide separates cash paid from credits consumed, works through an individual and a developer workload, and provides a reusable worksheet. Rates were checked on 7 November 2026; the examples are planning calculations, not tested invoices.

OpenRouter pricing at a glance

Billing route	Who bills for inference?	OpenRouter charge	What to watch
Free models	No paid model usage on eligible free routes	No credit-purchase fee without a purchase	Model availability and request limits
Standard credits	OpenRouter deducts provider list-price usage from credits	5.5% on credit purchases	A $0.80 minimum fee on standard card top-ups can make small purchases proportionally dearer
Business credits	OpenRouter deducts the same list-price inference rates	8% on credit purchases	Higher platform fee in exchange for Business controls and routing features
Standard or Business BYOK	Your provider bills inference directly	$0 on the first $25,000/month of list-price BYOK inference; 5% on the excess	Provider invoices, excess fees and any OpenRouter-funded fallback
Enterprise	Contract-dependent	Negotiated terms	Do not apply self-service rates to a custom agreement

OpenRouter states that it passes through provider list-price inference without a per-request markup. See OpenRouter’s official pricing and billing FAQ for the current plan terms. The figures below exclude VAT, other applicable taxes, exchange costs and optional paid services.

---

---

The five amounts that make up your real bill

A credit purchase is a prepayment, not proof that you’ve spent those credits. Treat the following as separate ledger entries:

1. Credits bought: the face value added to your OpenRouter balance. Buying $10 of credits does not mean you used $10 of inference.
2. Purchase fee: 5.5% on Standard or 8% on Business, calculated when credits are purchased. For Standard card top-ups, the stated $0.80 minimum fee applies.
3. Model consumption: fresh input, cached reads, separately priced cache writes, and output tokens multiplied by their respective model rates. On a normal credit-funded route, this amount reduces your balance.
4. BYOK provider invoice: model charges billed directly by the provider, potentially at negotiated rates. This is outside your OpenRouter credit purchases.
5. Additional OpenRouter usage: any BYOK overage fee and successful fallback requests funded from OpenRouter credits. These can exist even when your primary model uses BYOK.

One recurring source of billing confusion is assuming that a low OpenRouter credit deduction means the whole workload was cheap. A BYOK setup may show little credit consumption while a much larger bill accumulates at Anthropic, OpenAI or another provider. Reconcile both accounts before assessing cost.

Individual example: $10 of credits, but only $2.88 of inference

Assume 1,000 monthly conversations using openai/gpt-4.1-mini. Each averages 4,000 fresh input tokens and 800 output tokens, with no cache reads, paid tools or fallbacks. The model rates used here are $0.40 per million input tokens and $1.60 per million output tokens.

Item	Calculation	Amount
Fresh input	4 million tokens x $0.40	$1.60
Output	0.8 million tokens x $1.60	$1.28
Model inference consumed	$1.60 + $1.28	$2.88
Standard credits purchased	Face value	$10.00
Purchase fee	Greater of $0.55 or $0.80	$0.80
Cash paid today	$10 + $0.80	$10.80
Credits remaining	$10 – $2.88	$7.12

This shows why you shouldn’t describe the $10.80 checkout payment as the cost of 1,000 conversations. The workload consumed $2.88 of model credits, the purchase incurred a $0.80 fee, and $7.12 remains for future requests. If you are still choosing a model, our AI model comparison focuses on the cost per completed task rather than headline token rates alone.

Developer example: BYOK crosses the allowance, and a fallback draws credits

Now assume a high-volume coding service using anthropic/claude-sonnet-4.5 through its own provider key. This deliberately large monthly workload crosses the BYOK allowance. It assumes no provider discount, no extra tools, and successful cache reuse. Rates per million tokens are $3 for fresh input, $0.30 for cache reads, $3.75 for five-minute cache writes and $15 for output.

Token category	Monthly volume	List-price cost
Fresh input	4.7 billion	$14,100
Cached reads	3.2 billion	$960
Five-minute cache writes	100 million	$375
Output	1 billion	$15,000
Total model inference	8 billion input tokens and 1 billion output tokens	$30,435

Under these assumptions, the direct provider invoice is $30,435. Only $25,000 of list-price BYOK inference falls within OpenRouter’s no-fee allowance, leaving $5,435 subject to 5%: $271.75. You pay this fee in OpenRouter credits, not on the provider invoice.

Suppose the service also has $120 of successful OpenRouter-funded fallback usage during the month. It therefore consumes $391.75 of OpenRouter credits ($271.75 + $120). If the team buys $400 of credits on Standard, the purchase fee is $22, and the remaining credit balance is $8.25.

Cash and balance reconciliation	Amount
Direct provider invoice	$30,435.00
OpenRouter credits purchased	$400.00
Standard purchase fee	$22.00
Total cash paid, before tax	$30,857.00
OpenRouter credits consumed	$391.75
OpenRouter credits left	$8.25

The $30,857 cash figure includes $8.25 of unused credits. The calculated consumption plus this period’s credit-purchase fee is $30,848.75. Both totals are correct, but they answer different questions. Fallback costs are a separate planning allowance here, not evidence of failed attempts being charged twice. OpenRouter says failed model attempts are not billed for their model tokens; completed auxiliary services may still incur charges.

Reusable worksheet: calculate consumption, cash and credit balances separately

Copy these rows into a spreadsheet, with a separate column for each month or workload. Enter token prices in US dollars per million tokens. Keep provider discounts and fallback spending as editable inputs rather than burying them in your token rate.

Worksheet row	Calculation
Model inference at list price	(Fresh input x input rate + cache reads x read rate + cache writes x write rate + output x output rate) / 1,000,000
Direct provider invoice	BYOK model consumption adjusted for any actual provider discount; zero for a credit-funded route
BYOK fee	5% x MAX(0, monthly BYOK inference at OpenRouter list price – $25,000)
OpenRouter credits consumed	Credit-funded model usage + BYOK overage fee + paid OpenRouter fallback
Purchase fee	Credits bought x plan fee, subject to the applicable payment minimum
Credits remaining	Starting credits + credits bought – credits consumed
Total cash paid	Credits bought + purchase fee + direct provider invoice + applicable tax

Do not apply the 5% BYOK fee to a provider’s discounted invoice. The threshold and overage use OpenRouter’s corresponding list-price inference cost. Likewise, do not multiply every model request by 1.055 after already accounting for the Standard credit-purchase fee: that would double-count the platform charge.

Three ways a technically correct token estimate can miss the bill

Cache hits are a workload outcome, not a guaranteed discount

Repeated system prompts and tool schemas can qualify for cheaper cached reads. But the request needs to reuse a stable prefix and reach an endpoint holding the cache. A changed prompt prefix, expired cache, or provider switch can turn planned cached reads into full-price input. Some providers also charge extra to write the initial cache. For agents, record actual cache reads and writes by model and provider rather than assuming a fixed hit rate.

Fallback changes which account pays

Provider failover within the same model is different from switching to another model. Neither implies that an error itself costs a second full inference charge. But a successful response from a different, more expensive model can alter the final price. A BYOK configuration that permits OpenRouter-funded fallback can also start using prepaid credits when the own-key route is unavailable. Test both a successful BYOK path and a forced fallback before treating the billing route as fixed.

Your search, your sources

Make DIY AI a preferred source

See more of our reporting in Google Top Stories, AI Overviews and AI Mode.

Add DIY AI on Google

Automated tools can spend more than their chat interface suggests

Long coding sessions repeatedly send repository context, tool descriptions and prior results. An agent can also continue calling models after the user stops interacting with it. Set spend limits on individual API keys, review automatic top-ups, and ensure model access is restricted. BYOK workloads may be excluded from OpenRouter workspace budgets by default unless the relevant include-BYOK setting is enabled, so a low platform spend reading is not a complete safety limit.

For chat front-ends, the TypingMind review explains the separate cost of owning a client and paying for model APIs. For coding integrations, the OpenCode MCP setup guide covers how extra tools increase an agent’s context and operational complexity.

Should you use OpenRouter credits, BYOK or a direct API?

Use	Usually makes sense when	Main trade-off
OpenRouter credits	You want one balance, fast model switching and minimal provider account management	Purchase fee and less control over provider-level commercial agreements
OpenRouter BYOK	You already have provider keys or negotiated rates but want one routing interface	Two billing ledgers, BYOK allowance tracking and possible fallback credits
Direct provider API	You mainly use one provider, need direct billing and can maintain the integration	You own failover, observability and multi-provider migration work

For a hobby project, the engineering time saved by one API often outweighs a small credit fee. For a service concentrated on one provider with negotiated pricing, direct access can be cheaper. Compare the cost of completed tasks, not merely token price, and reconcile OpenRouter activity against provider invoices monthly.

OpenRouter pricing FAQs

Is OpenRouter free?

Some models have free routes with limits. Paid models consume credits or are billed by your own provider under BYOK. A free OpenRouter account does not make paid provider inference free.

Is the BYOK allowance still one million requests?

No. Older guidance described a monthly request-count allowance. The current Standard and Business allowance is $25,000 per month of inference measured at OpenRouter list prices, then 5% on the excess.

Does Business charge 8% on every API call?

No. Business charges 8% when you purchase credits. The published model inference rates remain the same, and BYOK has its separate allowance and overage rule.

Practical recommendation: start with one week’s real token usage, split cached reads and writes from fresh input, and keep the provider invoice separate from credit purchases. That is the shortest route to a forecast you can check against actual billing.