OpenRouter

OpenRouter is a gateway, not a model lab. One key reaches 600+ models from 80+ providers at each provider's list price. Its fee is 5.5 percent when you buy credits on Standard, or 8 percent on Business, never per request. Data terms are set per provider, and you can filter routing by them.

Facts at a glance

Prices checked 2026-10-31. We do not list model prices here because they belong to the vendors behind the gateway, and each has its own entry: Anthropic, OpenAI, Google AI, DeepSeek, Groq and others.

How the money works

OpenRouter's FAQ is blunt about it. Input and output tokens are billed per model at posted rates. Each request deducts its cost from your credit balance, and the platform does not mark up inference. The fee is charged at the moment you buy credits.

A quick sum shows what the fee costs in real money. Buy $1,000 of credits on Standard and the fee is $55. On Business it is $80. Neither fee changes the per-token price you pay. The gap between the plans is therefore a paid-for-features gap, not a price gap, and for most small teams Standard is the right tier.

Bring your own key. With BYOK the provider bills you directly. Standard and Business include $25,000 of list-price inference a month through BYOK with no OpenRouter fee, and 5 percent applies after that. Enterprise allowances are custom. If you already have an Anthropic or OpenAI contract with good limits, BYOK lets you keep the contract and still use OpenRouter's routing.

Failed requests. The FAQ says an attempt that errors is not charged for its model tokens. With fallback enabled you pay only for the run that answers. A feature called Zero Completion Insurance extends the same idea to a response that returns nothing. It applies automatically to all accounts when a response has zero completion tokens and a blank finish reason, or ends with an error finish reason, and it covers prompt, completion and reasoning tokens. It does not cover auxiliary services such as web search or file parsing that already ran.

Rate limits and credits

OpenRouter says capacity is governed globally. Making extra accounts or API keys will not raise your limits. Different models carry different limits, so spreading load across models is the lever it names. You can check the limit and the credit left on a key with a GET request to https://openrouter.ai/api/v1/key, which returns the remaining credit, usage today and all time, and whether the account has ever bought credits.

Free model variants, whose ids end in :free, have their own per-minute and per-day request caps, which depend on how many credits you have ever purchased. The limits page gives 20 requests per minute either way, and 50 requests per day if you have bought fewer than 10 credits in total or 1,000 per day if you have bought at least 10. Requests that hit a platform limit return HTTP 429. Cloudflare DDoS protection will also block traffic that dramatically exceeds reasonable usage.

Data policy: two layers

This is where a gateway is different from a lab. There are two parties handling your prompt, and each has its own rules.

1. OpenRouter itself. Its own handling of your prompts is described in its privacy policy. Its FAQ says prompts and completions are not logged by default unless you opt in to logging, and the privacy policy says providers' retention practices vary. Your account setting about training does not change this.
2. The provider that serves the model. Each provider has its own logging, retention and training policy. OpenRouter reflects these as structured data on each endpoint and shows them on a documentation page. It does not change routing based on retention, so you have to filter.

You have three controls.

The regional model list is available by calling /api/v1/models/user through the regional hostname, so you can see which models qualify. Server tools and plugins that call third-party services, such as search engines and sandboxes, follow those services' policies, not the model provider's.

The sensible reading is that OpenRouter gives you the knobs, not the answers. A free-plan user who sends private code through the default routing has accepted whatever the cheapest eligible provider's terms are. A Standard user who turns on data policy filtering has not. Check the setting before you point an agent at a work repository.

Routing, fallbacks and what a gateway changes

OpenRouter's pitch, in its own FAQ wording, is a unified API and fallbacks, so that you get better uptime. In practice a gateway changes five things about how you buy and run models.

The last row is the one that decides most enterprise reviews. A direct contract with Anthropic or OpenAI gives you one set of terms to read. A gateway gives you two layers, and the second layer can change from request to request when a fallback sends your prompt to a different provider. That is fine for a public data set and a problem for regulated data. The fix is a filter, not a policy document: use data policy filtering and pin providers for the models you rely on.

Auto-routing and preferred vendor selection are on Standard and above, so the free plan gives you a single, manual choice. Budgets and spend controls are also Standard and above. That matters for agents. A loop that goes wrong on the free plan can only burn the free quota, while on Standard a runaway agent can spend your whole credit balance unless you set a budget on the key.

Choosing a plan, scenario by scenario

The fee arithmetic is worth running for your own spend. At $500 a month of credit purchases, Standard costs $27.50 and Business $40. At $5,000 a month, Standard is $275 and Business $400. The Business premium of 2.5 points buys regional routing and notifications, so the question is whether those two features are worth a flat share of your spend. For a team that needs the EU hostname, they usually are. For one that does not, they are not.

What a gateway does not do

OpenRouter does not make a model better, cheaper per token, or more private. It will not add a zero-retention guarantee to a provider that does not offer one. It cannot give you a feature that a provider has not exposed, so provider-specific tools, such as a vendor's code execution, may be missing or behave differently through the compatible API. It also adds one more hop, with its own availability. If OpenRouter is down, so are your fallbacks, unless you built a second path.

Two checks before you commit. First, test the specific feature you need, such as tool calls, structured output or prompt caching, on the exact model through the gateway, because gateway support varies by provider. Second, compare the per-token price on OpenRouter with the vendor's own page for your top model. The FAQ says the prices match list prices, which means the real extra cost is the credit fee plus the extra hop.

Moving a project onto OpenRouter

1. Start with one model. Point an existing OpenAI-compatible client at https://openrouter.ai/api/v1 and use a model id of the form provider/model.
2. Add the data filter before any real data goes through.
3. Set a budget on the key, then add a second model as a fallback.
4. Compare cost against direct billing for a week of your own traffic, including the credit fee.
5. Decide on BYOK if you hold vendor contracts, so your negotiated limits still apply.

Moving off is as easy as the move on. Because the interface follows the OpenAI shape, a switch back to a direct vendor is usually a base URL and key change, plus any gateway-only features you adopted. Teams that adopted auto-routing should write down which model answers which task before leaving, because direct billing will not choose for them.

Which agents can use it, and how

OpenRouter documents a Claude Code route. The token goes in ANTHROPIC_AUTH_TOKEN, because Claude Code sends that as a bearer token, and ANTHROPIC_API_KEY must be blank so the two do not conflict:

export OPENROUTER_API_KEY="<your-openrouter-api-key>"
export ANTHROPIC_BASE_URL="https://openrouter.ai/api"
export ANTHROPIC_AUTH_TOKEN="$OPENROUTER_API_KEY"
export ANTHROPIC_API_KEY=""

OpenRouter warns that the order matters. If the key is defined after the line that expands it, or injected by a secrets manager later, the token is empty. See Claude Code.

Aider has a native route:

export OPENROUTER_API_KEY=<key>
aider --model openrouter/<provider>/<model>

OpenCode, Cline and Roo-style agents accept OpenRouter as an OpenAI-compatible provider with the base URL https://openrouter.ai/api/v1. See Aider, OpenCode and Cline. For a gateway, the useful trick is a single key that lets you try a new model the day it appears, without a new account.

Alternatives

OpenRouter has several rivals, with different trade-offs, all covered on OpenRouter alternatives.

Who should pick OpenRouter

A pre-launch checklist

1. Set the training setting to your policy, for paid and free models separately.
2. Turn on data policy filtering if the data is sensitive.
3. Use eu.openrouter.ai or us.openrouter.ai if location matters, and confirm the model is on the regional list.
4. Set a budget and spend control on the key.
5. Check /api/v1/key for the remaining credit before a big job.
6. Pin a provider for models where retention differs between hosts.
7. Log which provider answered each request, so a data review can find it later.

Our take

OpenRouter is a convenience layer and it is honest about the price of convenience: a fee on credits and no markup on tokens. It earns its place when you want to try models quickly or need fallbacks, and the privacy controls are real but opt-in. The risk is that the gateway makes it easy to forget who is actually handling a prompt. Use the data policy filter and in-region routing on any work that matters, and remember that the 8 percent Business fee is what the regional hostnames cost you. For a single-vendor workload with a large volume, buying direct from the lab is simpler and cheaper.

We have not run tests for this page. It reflects OpenRouter's published pages on 2026-10-31. OpenRouter does not pay for placement here.

Related pages

OpenRouter alternatives, LLM API prices, Claude Code, AI hosting, Hugging Face Inference.