OpenRouter Pricing: Request Cost, Credit Fees, and BYOK

To budget for OpenRouter, separate inference usage from the money deposited into your balance. If you bring a provider key, calculate BYOK separately. A single per-million-token rate does not describe all three flows.

Model rates and credit purchases are different lines

According to the official FAQ, requests are deducted from credits and prices depend on the model and provider. OpenRouter says it passes through provider inference rates without markup, while charging a fee when credits are purchased. On September 11, 2026, the pricing page lists 5.5% for pay-as-you-go.

Before paying, compare credits received, the fee shown, and the total charge. Minimum fees, taxes, and payment-method conditions must be checked in the live checkout. A conditional example: if $100 in credits has only a 5.5% fee added, the payment is $105.50. This excludes taxes, currency conversion, and other charges. After using $20, $80 in credits remains; the full deposit is not the cost of those first requests.

Calculate the actual workload

For simple text use without caching or extra paid features:

request_cost = input_tokens / 1000000 * input_price
             + output_tokens / 1000000 * output_price

Take the real BetterToken price for glm-5.3 in the GLM group, checked in its current pricing catalog on September 11, 2026. Input is $1.232 and output $3.872 per million tokens. For 8,000 uncached input and 2,000 output tokens:

input_cost = 8000 / 1000000 * 1.232 = $0.009856
output_cost = 2000 / 1000000 * 3.872 = $0.007744
total_cost = $0.0176
1000_requests = $17.60

BetterToken adds no separate fee to this usage cost. The result remains $0.0176 per request or $17.60 for one thousand identical requests; do not multiply it by 1.055 because OpenRouter’s fee does not apply to BetterToken. The request sizes are illustrative, the rates are current, and cache tokens or additional calls are excluded. This is a BetterToken calculation, not a comparison with the current OpenRouter rate for GLM-5.3.

Use your real request lengths, including conversation history, and separate short from long tasks. For an agent, include every call needed before acceptance, not only the final answer. Caching, images, reasoning, and per-request prices can change the formula.

OpenRouter usage is available in Activity. For BetterToken, use the pricing page and your Workspace, which shows the model, tokens, and charge. Compare the same accepted task and the same billing units.

A free model does not mean unlimited workload

The OpenRouter pricing page shows 50 requests per day for the free plan. This does not promise identical conditions for every account or paid route. The limits documentation separates spend limits from request rate limits: a 402 calls for checking balance and key limits, while a 429 may come from the platform or an upstream provider.

For a 429, inspect the reason and honor Retry-After when present. Avoid infinite retry loops. If continuation switches to a paid model, include that cost before the run. Free access can test compatibility, but critical work needs an explicit response to refusal or exhausted limits.

BYOK: check both bills

BYOK means sending requests through OpenRouter using your own provider key. The current pricing page says pay-as-you-go includes up to $25,000 per month in list-price inference without a BYOK fee, followed by 5%. This amount measures the allowance; it is not free requests and does not pay the provider bill.

In the BYOK settings, fallback to OpenRouter shared capacity can consume OpenRouter credits. Check the provider bill, BYOK fee, and requests that used the normal balance. Review key order and shared-capacity fallback before forecasting costs.

If one direct API is enough, compare its final bill for the same work. If multiple models matter, include the value of one interface and consolidated records. A small, bounded batch is enough for the first comparison: cost per accepted result is more useful than input-token price alone.