What is OpenRouter? One key, 400+ models, and what it costs you

What OpenRouter actually is

OpenRouter sits between your code and every model vendor. You write one integration against an OpenAI-shaped API, hold one key, top up one balance, and get access to Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot, Mistral, Qwen, and a long tail of open-weight models served by specialist inference hosts. That is the entire product idea, and it is a good one: the alternative is five vendor accounts, five billing relationships, five SDKs, and a procurement conversation each time a new model matters.

It is a hosted gateway, not a library. There is nothing to deploy and nothing to operate, which is the main axis it differs on from LiteLLM.

The numbers, checked this month

Vendor homepages round up. These come from OpenRouter's own public API, read on 19 August 2026.

How billing works

This is the part people get wrong, in both directions. OpenRouter does not take a cut of inference. Its FAQ is explicit: "We pass through the pricing of the underlying providers without any markup, so you pay the same rate as you would directly with the provider."

The fee is on the top-up. You buy credits, the fee is added on top of the purchase, and the credits then spend at provider list price.

So a $100 top-up costs $105.50 on a card and buys $100 of tokens. The effective overhead against going direct is 5.5%, and it is flat: it does not scale with how expensive the model is, and it does not compound.

Free models and their real limits

Seventeen model ids end in :free today, mostly NVIDIA Nemotron variants, Google Gemma 4, Z.ai GLM 5.2, OpenAI gpt-oss-20b, and a few preview models. They are genuinely free of token charges. They are not free of limits.

The $10 threshold is a lifetime purchase, not a balance, so it survives spending the credits. OpenRouter's own docs say free models are "usually not suitable for production use", which is the right way to read a 20 RPM ceiling: they are for evaluation, hobby projects, and fallback tiers, not for a coding agent that fires ten tool-call turns a minute.

BYOK: your keys, their routing

Bring-your-own-key lets you attach your existing Azure AI Foundry, Azure OpenAI, Amazon Bedrock, Google Vertex, or OpenAI credentials and keep your negotiated rates, committed spend, and provider-side rate limits, while still using OpenRouter's single API and its analytics.

• The fee is 5% of what the same model and provider would have cost at list price, deducted from your OpenRouter credit balance.
• It only applies above $25,000 of list-price inference per month on pay-as-you-go, or $200,000 on enterprise. Below that, BYOK is free.
• BYOK endpoints are always tried first and bypass your provider ordering.
• BYOK spend does not count toward guardrail or workspace budgets by default. There is a toggle.
• A BYOK key cannot override a zero-data-retention restriction. Data policy still applies.

Routing, and why the price you see is a floor

A model on OpenRouter is usually served by several providers at several prices, and the number on the model card is the cheapest endpoint. That is not the same as the price you will pay, because routing depends on availability, your data-policy filters, and your sort preference.

Here is the endpoint list for one model, read from /api/v1/models/openai/gpt-5.6-sol/endpoints on 19 August 2026.

The card shows $2.50/$15. OpenAI's own published rate for that model is $5.00 input and $30.00 output at short context. Nothing here is a markup, and nothing is hidden: it is a genuine spread across endpoints, and OpenRouter passes each one through. But "cheapest listed" is a routing outcome you have to ask for, not a price you are promised.

{
  "model": "openai/gpt-5.6-sol",
  "provider": {
    "sort": "price",
    "only": ["openai"],
    "allow_fallbacks": false,
    "data_collection": "deny"
  }
}

By default OpenRouter load-balances on price weighted by the inverse square of cost, with uptime taken into account: its docs give the example that a $1/M provider is nine times more likely to be picked than a $3/M one. sort: "price" or the :floor suffix drops the balancing and always takes the cheapest; sort: "throughput" or :nitro takes the fastest. allow_fallbacks: false pins you to the top provider and fails rather than silently spending more.

Privacy, logging, and attribution

OpenRouter does zero logging of prompts and completions unless you opt in, and opting in earns a 1% discount on usage. Separately, your account settings carry a training toggle for paid and free models: with training opt-out on, OpenRouter will not route to providers that train on your data. Each provider's own retention shows in a table as zero retention, retained for N days, or unknown, and you can filter per request with data_collection: "deny". Enterprise accounts can pin processing to a region via https://eu.openrouter.ai or https://us.openrouter.ai.

App attribution

Send an HTTP-Referer header and your app gets a page on openrouter.ai/rankings with daily, weekly, and monthly token usage; X-OpenRouter-Title sets the display name, and X-OpenRouter-Categories tags it (two per request, ten per app). The referer header is required: without it no app page is created. It is free distribution for a developer tool, and it is also public disclosure of your token volume, which some teams will not want.

When OpenRouter is the right call, and when it is not

Use it when

• You want to try models faster than you can open vendor accounts. This is the strongest case and it is not close.
• You are shipping a product that lets users pick a model. One integration, 400+ options.
• You want failover across providers for the same model without writing it yourself.
• You are a solo developer or small team and the 5.5% is cheaper than the hours of managing five billing relationships.

Do not use it when

• You call one model, all day, at volume. 5.5% of a large bill buys nothing you cannot get from the vendor SDK.
• Latency is your product. There is an extra hop, and the vendor does not stop existing.
• You need a provider-specific feature on day one. Aggregators normalize, and normalization lags: prompt caching semantics, extended thinking controls, and beta headers are where the edges show.
• Your compliance review counts data processors. One more name on the list can cost more than the fee saves.

Questions people ask

Keep reading

Continue through LLM gateways and routing with OpenRouter pricing: the 5.5% fee, free models, and what routing costs.