OpenRouter
OpenRouter is a unified API gateway that routes requests to more than 400 large language models from over 70 providers through a single OpenAI-compatible endpoint, using prepaid credits with automatic failover and no markup on inference.
What is OpenRouter?
OpenRouter is an aggregator, or gateway, sitting between your code and the large language model market. Its own Terms of Service describe it plainly as a large language model aggregator through which users reach third-party APIs to use a variety of generative AI models listed on the site. Instead of holding an account, a contract and a separate integration with every model vendor, you hold one of each with OpenRouter and reach the rest through a single OpenAI-compatible endpoint.
The catalogue is the main asset, and it is verifiable rather than merely advertised: the first-party models endpoint returns exactly 400 models from 58 different publishers, including 14 free variants. OpenAI accounts for 95 of them, Qwen 49, Google 39, Anthropic 28 and Mistral 18, with the longest context window in the catalogue reaching two million tokens. Roughly 150 are text-to-text; the rest are multimodal, with more than a hundred accepting text, images and files, and a smaller group handling audio and video or returning images and audio. A companion providers page lists 80 inference providers and, unusually, publishes for each one whether it trains on prompts, how long it retains them, whether it supports bring-your-own-key, and where it is headquartered.
Around that catalogue sits the routing layer that justifies the product. Requests can be steered to the lowest price, the highest throughput or the most available provider, and OpenRouter falls back automatically to the next provider when one errors, which is the failure mode most teams otherwise handle with their own retry code. Model variants let you append behaviour to a request: a free tier, an extended context, reasoning by default, attached web results, faster throughput or lowest price.
Commercially it works on prepaid credits rather than a subscription. A permanent free tier gives 25 or more free models at 50 requests a day; paid use adds a 5.5% platform fee on credit purchases, while inference is billed at the provider's own rate, since OpenRouter states it never marks up provider pricing. Enterprise accounts add SSO/SAML, SCIM, admin controls, contractual SLAs, invoiced billing, one-click zero data retention and in-region routing for data residency.
What it does
- Call more than 400 models through one OpenAI-compatible endpoint
- Route each request automatically to the cheapest, fastest or most available provider
- Fail over to another provider when one errors or goes down, without changing any code
- Compare models on price, latency and throughput before committing to one
- Cap spending with per-key credit limits, scheduled resets and workspace budgets
- Enforce a data policy, including one-click zero data retention
- Broadcast request traces to Langfuse, Datadog, Braintrust, Weave or S3
When to use OpenRouter / When not to
A quick filter to help you decide if OpenRouter is the right fit.
When to use OpenRouter
- Back-end and full-stack developers who call several different LLMs from a single OpenAI-compatible SDK
- Platform and LLMOps teams that need provider failover, per-key budgets, SSO/SAML and unified request reporting
- Founders and indie hackers — the pricing page addresses indie hackers, AI-native startups and enterprises by name
- Product managers who have to choose the model inside their product, using public rankings, evaluations and price-latency comparison
- Researchers, journalists and analysts working from the aggregate usage dataset and the Datasets API
When not to use OpenRouter
- Non-technical users looking for a ready-made consumer assistant: an account, prepaid credits and an API key are all required first
- Anyone who needs a first-party mobile app, or an interface in any language other than English
- Production workloads running on the free models, which carry low rate limits and are described as unsuitable for production
- Teams that need a predictable fixed monthly bill rather than per-token usage billing
- Anyone wanting to fine-tune or train their own models — OpenRouter routes inference, it does not train
How to use OpenRouter
A typical end-to-end flow, from setup to results.
- Create an account — an individual one, or an organisation account if a team will share it
- Buy credits on the credits page: the minimum is USD 5.00 per transaction and the maximum USD 25,000
- Generate an API key from the keys page in settings
- Point your existing OpenAI SDK at the OpenRouter endpoint, or call the chat completions endpoint over plain HTTP
- Pick a model from the catalogue, or let automatic routing choose, appending a variant such as lowest-price, faster-throughput or web-results when useful
- Set your data policy in the privacy settings: optional prompt logging, zero data retention, and which providers are allowed
- Cap spending with per-key credit limits, daily, weekly or monthly resets, auto recharge and workspace budgets
- Track consumption on the activity dashboard, or ask for usage data to be returned inside each API response
- Connect observability by broadcasting traces to Langfuse, Datadog, Braintrust, Weave or S3
- Try models without writing code in the chat room, or contact sales for SSO, SLAs and invoiced billing
Pros & Cons
Pros
- One integration and one contract replace five or more separate provider relationships
- No markup on inference — OpenRouter states that its pricing always matches provider pricing
- No subscription and no long-term contract required: credits are bought as needed
- A permanent free tier with 25 or more free models, 4 providers and 50 requests a day
- Unusual transparency about providers: training, retention, bring-your-own-key support and headquarters are published for 80 of them
- Prompts and completions are not logged by default, and zero-retention routing is available
- OpenAI compatibility makes migrating an existing integration close to costless
Cons
- A 5.5% platform fee applies to credit purchases on top of inference pricing
- Costs are variable by design: priced per model and per token type, with no fixed monthly plan
- No volume discounts are currently offered
- Credits may expire 365 days after purchase, refunds are limited to 24 hours, platform fees are never refundable and crypto payments never are either
- Output ownership is not guaranteed by OpenRouter: it depends on each model's own terms
- Privacy depends on the provider you route to — 2 of the 80 listed train on prompts and 23 retain them
- Free models carry low rate limits and are explicitly described as unsuitable for production
Pricing & Plans
A permanent free tier is available at no cost, covering 25 or more free models across 4 providers, limited to 50 requests per day with community support. Paid use is not a subscription: it runs on prepaid credits, with a minimum purchase of USD 5.00 per transaction and a maximum of USD 25,000. On the Pay-as-you-go tier a platform fee of 5.5% applies to credit purchases, while inference itself is charged at each provider's own rate with no markup, priced per model and varying between prompt, completion, image and reasoning tokens. There is no minimum spend and no commitment. Enterprise terms are quoted individually and may include fee discounts and volume commitments.
- 25+ free models
- 4 free providers
- 50 requests per day
- chat and API access
- activity logs and export
- auto-routing
- budgets and spend controls
- prompt caching
- 400+ models
- 70+ providers
- high global rate limits
- no minimum spend
- management API key
- BYOK covering USD 25
- 000 of list-price inference per month with no fees then 5%
- payment by credit card and crypto
- 400+ models and 70+ providers
- admin controls
- managed policy enforcement
- SSO/SAML
- contractual SLAs
- optional dedicated rate limits
- volume commitments
- BYOK covering USD 200
Data, GDPR & hosting
A consolidated view of how OpenRouter handles your data.
GDPR overview
GDPR is addressed explicitly rather than implied. The enterprise page presents OpenRouter as a single GDPR-compatible, SOC-2 compliant partner with SLAs, and offers EU region locking. A Data Processing Agreement is published and referenced from clause 10.2 of the Terms. The privacy policy, last updated on 6 July 2026, states that personal data may be transferred to servers in the United States or outside the EEA and the UK, and relies on Article 45 adequacy decisions and Article 46 standard contractual clauses. It sets out a legal-basis table covering contract, legitimate interest, consent and legal obligation, and lists rights of access, portability, erasure, rectification, objection, restriction and withdrawal of consent, exercised by email. Two gaps are worth noting: no Article 27 EU representative and no Data Protection Officer is named anywhere on the site.
Who owns the data?
Under the Terms of Service you keep the copyright and any other proprietary rights you hold in the Input you send. Ownership of the Output is deliberately not settled by OpenRouter: it is governed by the Model Terms of whichever model produced it. By using the service you grant OpenRouter a non-exclusive, transferable, worldwide, royalty-free and sublicensable licence to host, cache, store, reproduce, transmit, display and reformat your content, solely to operate and provide the service. OpenRouter states it does not use Inputs or Outputs for model training, and that it has opted out of training with the models it uses where possible. Individual model providers may nonetheless retain and train on what you send them.
Reuse rights
You can reuse your own Inputs freely: the copyright stays with you and OpenRouter claims no ownership of them. Outputs are the point that needs checking, because OpenRouter does not grant those rights itself — each model's own Model Terms decide what you may do with what it generates, so commercial reuse has to be verified model by model rather than settled once for the whole platform. You remain responsible for holding the rights to everything you submit. Content you deliberately publish through the sharing feature can be viewed and copied by everyone you give access to, and OpenRouter reserves the right to remove or block Inputs it judges to breach the Terms. Opting into prompt logging widens the licence considerably, so it is a choice worth making knowingly.
Data retention & training
Hosting summary
The privacy policy states that personal data may be transferred to OpenRouter's servers in the United States, or to other countries outside the European Economic Area and the UK, with transfers framed by Article 45 adequacy decisions and Article 46 standard contractual clauses. Data provided to OpenRouter is described as stored on secure servers behind firewalls. Enterprise customers can narrow this: the platform advertises zero data retention, GDPR compliance and EU region locking, plus a Sovereign AI capability offering in-region routing and data residency controls that keep workloads inside national or regional boundaries. The effective processing jurisdiction, however, follows the provider a request is routed to, and the providers page shows headquarters spread across the United States, Singapore, China, Sweden, Spain, the Netherlands and Israel. One technical caveat: the domain resolves to a Cloudflare anycast address, which is a content-delivery point of presence and not a storage location. Image, audio and video files are not kept beyond the time needed to route the request, except for abuse detection, security, billing or legal compliance. Governing law is New York.
Where OpenRouter works
Country-level availability.
Not available in
Things to keep in mind
Risks and trade-offs to weigh before adopting OpenRouter.
- Concentration risk: your entire LLM consumption ends up behind one key and one contract
- Output rights vary from model to model, so any commercial use means reading the terms of each model you actually call
- Opting into prompt logging to get a discount switches you to a perpetual, irrevocable licence that includes licensing or selling that content in anonymised form
- Some providers train on or retain prompts, so a routing choice is also a privacy choice — and it is easy to make it by accident
- Voice or face data sent to cloning models counts as biometric data under GDPR, the CCPA and Illinois BIPA, and you are responsible for the consent of the people it depicts
- Usage-based billing means a badly bounded agent loop can consume a lot of credit very quickly if per-key limits are not set first
- Disputes go to mandatory arbitration under New York law, with class actions waived
Setup & Integrations
Technical difficulty
Low for a developer, and not really applicable to anyone else. Without code, the chat room lets you try models right after signing up. With code, setup is three steps — account, credits, API key — and the API is a drop-in OpenAI replacement, so most existing SDKs work unchanged. The prerequisites are modest but real: calling an HTTP API, handling a secret key, and understanding per-token billing well enough to set spend limits. Provider selection, data policies and zero retention are optional. Only the enterprise features add work, since SSO/SAML and SCIM need directory configuration.
Deployment
Integrations
Supported languages
Behind OpenRouter
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Does OpenRouter take a margin on model pricing?
How many models and providers are actually available?
Is there a free option?
What is the cheapest way to start paying?
Are my prompts logged?
Does OpenRouter train models on my data?
Can I prevent any retention or training entirely?
Do I have to rewrite my code to switch?
What happens when a provider goes down?
Are credits refundable?
Should you pick OpenRouter?
OpenRouter solves a problem that only appeared once teams started using more than one model: the cost, in contracts and in code, of holding a relationship with every vendor. Its answer is credible and, more importantly, checkable. The advertised catalogue matches what the public models endpoint actually returns, and the providers page goes further than most vendors would by publishing, provider by provider, who trains on prompts, who retains them and where they are based. The commercial story is equally legible: no subscription, prepaid credits, a stated 5.5% platform fee and a repeated commitment never to mark up inference.
The reservations follow from the same architecture. OpenRouter is an intermediary, so two things it cannot give you are precisely the two that matter most in a contract. Ownership of what a model generates is set by that model's own terms, not by OpenRouter, which means commercial reuse has to be checked model by model. And the privacy guarantee you actually get depends on where your request is routed: the platform itself does not train on your data and does not log prompts by default, but a fifth of the listed providers retain them. Zero data retention and policy-based routing exist to close that gap, and they are worth configuring on day one rather than later.
Financially, expect variability rather than a line item: no volume discounts, credits that may expire after a year, and a 24-hour refund window. The platform looks mature — GDPR and SOC-2 are claimed, a DPA is published, data residency is available — and it is plainly aimed at developers rather than end users. For a team already calling two or more models, it is a straightforwardly sensible consolidation. For anyone wanting a finished assistant, it is the wrong layer of the stack.
- Choosing a selection results in a full page refresh.
- Opens in a new window.