OpenRouter
OpenRouter is the widest managed gateway to LLMs: 500+ models across 80+ providers behind one OpenAI-compatible endpoint, with automatic fallback, MCP support, and no per-token markup. Its pricing is honest about the token pass-through but less forthcoming about the $0.80 minimum fee that makes small top-ups cost 8–16% effectively, and about the fact that EU/US in-region routing requires the 8% Business tier. BYOK is free up to $25,000/month, which is a non-issue for almost everyone. It is a strong default for model breadth and zero-ops access—not for self-hosting, small-budget cost optimization, or contractual uptime guarantees.
What is OpenRouter?
OpenRouter is a managed LLM gateway, not a model provider. Its official framing is “build with every model, from one platform”: a single OpenAI-compatible API endpoint at openrouter.ai/api/v1 that fronts 500+ models from 80+ providers, with automatic fallback when a provider fails. The product surface has expanded beyond the API to include typed SDKs in TypeScript, Python, and Go, an agent framework (@openrouter/agent), a CLI for local telemetry and replay, and a hosted MCP server at mcp.openrouter.ai/mcp that gives any MCP client—Claude Code, Cursor, or a custom agent—access to the full model catalog.
The gateway's core value proposition is breadth without operational burden. A developer can start with a free model, move to a paid one by changing a model string, and add fallbacks so that a provider outage does not become their outage. The same code works across OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and dozens of others without provider-specific SDKs. The trade-off is dependency: OpenRouter is the single point through which every request passes, and it is not self-hostable. Teams that need VPC-level control, data residency outside EU/US, or contractual uptime guarantees on self-serve tiers must look at LiteLLM or Portkey instead.
Pricing is where OpenRouter's honesty and its fine print diverge. The company is correct that it does not mark up tokens: every model bills at the provider's published list rate, and independent checks confirm the catalog matches provider pricing. The revenue comes from the credit purchase fee: 5.5% by card on Standard, 5% for crypto, 8% on Business. The $0.80 minimum on card fees is the number that changes the math for small users—a $5 top-up pays 16% effectively, not 5.5%. BYOK (bring your own provider keys) is free up to $25,000/month of list-price inference on Standard, then 5% of the excess, with Enterprise defaulting to $200,000. For almost every user, the BYOK fee is irrelevant; the complexity of key management is the real cost.
Data handling is more privacy-conscious than the gateway category's reputation might suggest, but it requires reading two documents to understand fully. OpenRouter does not store prompts or responses by default. There are two opt-ins: private input/output logging (off by default, data stored for the user's own review and not accessed by OpenRouter) and OpenRouter's use of inputs/outputs for product improvement (off by default, offers a 1% discount, not model training). CSAM screening can trigger retention regardless of settings. Metadata (token counts, latency) is always stored to power reporting and rankings. Enterprise adds ZDR routing with contractual guarantees and org-wide policies. The 1% discount framing is the one place where a pricing incentive and a data decision are presented together, and users should read the boundary before accepting it.
How does OpenRouter work?
A developer creates an API key from the dashboard and points an OpenAI-compatible client at openrouter.ai/api/v1. The model string follows an author/slug format (e.g., anthropic/claude-sonnet-4). If the client is already using the OpenAI SDK, the change is limited to the base URL and the key. Requests are routed to the selected provider; if that provider fails or is rate-limited, OpenRouter attempts the next provider in the fallback order. Provider ordering, allow/deny lists, and fallback behavior are configurable per request. The response is returned in the same shape as the provider's native API, with OpenRouter metadata attached.
For BYOK users, provider keys are encrypted and stored in workspace settings. The BYOK routing model has three modes: prioritized (tried before OpenRouter's shared capacity), fallback (tried only after shared capacity), and shared-capacity-default (try the user's key first, then fall back). Two “never use shared capacity” modes let a user force all traffic for a provider through their own keys, accepting rate-limit errors rather than spending OpenRouter credits. Provider ordering interacts with BYOK: providers named in the order array are tried first, and within each provider the user's BYOK key is tried before OpenRouter's shared capacity.
The MCP server at mcp.openrouter.ai/mcp exposes OpenRouter's model catalog and chat functionality as MCP tools, so an MCP client can discover models, run completions, and route across providers without a custom integration. It supports OAuth-based connection from Claude Code, Cursor, and other clients. The CLI captures SDK telemetry locally and replays requests in a web viewer, which is useful for debugging agent loops without adding a third-party observability vendor.
Use Cases
- Prototyping across multiple model families without maintaining separate API keys, SDKs, or billing relationships
- Production agents that need automatic fallback when a primary provider fails or rate-limits, configured per request or per provider
- Accessing a broad catalog of models—including open-weight models not available through first-party APIs—through one OpenAI-compatible interface
- Giving an MCP client access to 500+ models through a single hosted endpoint without building a custom model router
- Teams that want per-key budgets, spend controls, and workspace separation without building a gateway in-house
- European or US teams that need in-region routing on Business, or ZDR routing with contractual guarantees on Enterprise
- Free-model experimentation and light development, with the option to raise the daily limit by making a one-time $10 credit purchase
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Free | $0 | per month | Access to 25+ free models (IDs ending in :free) across 4 free providers. 20 requests per minute and 50 requests per day. Rate limits apply at the account level regardless of purchase history. No prompt caching, no budgets or spend controls, no BYOK, no in-region routing. 5 workspaces. No card required. | — |
| Standard (pay-as-you-go) | 5.5% on credit purchases | per top-up ($0.80 minimum) | 500+ models across 80+ providers at provider list price. 5.5% platform fee on card purchases, $0.80 minimum; 5% for crypto (USDC). $5 minimum top-up, $25,000 maximum per transaction. BYOK free up to $25,000/month of list-price inference, then 5% of the excess. Prompt caching, budgets and spend controls, provider passthrough rate limits. No in-region routing on this tier. 5 workspaces. | — |
| Business | 8% on credit purchases | per top-up | Everything in Standard plus EU and US in-region routing (eu.openrouter.ai / us.openrouter.ai). 8% platform fee on credit purchases. 1,000 workspaces. No contract or minimum spend required; upgrade in settings and downgrade 30 days later. | — |
| Enterprise | Custom | annual | Custom fee discounts, contractual SLAs, SSO/SAML, SCIM, dedicated support, org-wide policies, ZDR routing with contractual guarantees, and custom BYOK allowances (default $200,000/month of list-price inference before the 5% fee). Billed via invoice or purchase order. | — |
Common Limitations
- The $0.80 minimum fee makes small top-ups disproportionately expensive · OpenRouter's 5.5% platform fee has a $0.80 floor on card purchases. That floor changes the effective fee depending on top-up size: a $5 top-up pays $0.80, which is 16% of the amount loaded, not 5.5%. A $10 top-up pays 8%. The fee does not approach the advertised 5.5% until roughly $15 is loaded at once. For an indie developer testing a project with a $5 deposit, the real cost of credits is 16% above the token list price before a single request is made. Crypto top-ups avoid the card floor but charge a flat 5% with no minimum, which is cheaper for small amounts and slightly cheaper for large ones. The minimum is documented, but it is not the number most users anchor on when they read “5.5%,” and it means the gateway's cost advantage over going direct is smaller than the headline suggests for anyone who tops up in small increments.
- BYOK's $25,000/month free allowance is a paper benefit for almost everyone · OpenRouter's BYOK offer is structured as a free allowance measured in list-price inference dollars: $25,000 per month on Standard before the 5% fee applies. For the overwhelming majority of developers and small teams, monthly inference spend is measured in tens or hundreds of dollars, not tens of thousands. The $25,000 threshold means BYOK is functionally free for almost all users—which is genuinely useful—but the framing creates a misleading impression that BYOK is a premium feature with a metered tail. In practice, BYOK's real trade-off is not the 5% fee (which most users will never trigger) but the operational complexity of managing provider keys, understanding key priority and fallback order, and accepting that requests routed through your keys bypass OpenRouter's rate-limit management. The fee is a non-issue below $25K; the complexity is the actual cost.
- Free-model limits are request-count caps, not token caps, and the 50/day ceiling is a hard stop · Free models on OpenRouter are limited to 20 requests per minute and 50 requests per day. The daily limit rises to 1,000 only after a one-time purchase of at least $10 in credits—not a subscription, but a lifetime purchase threshold. Until that purchase, a developer building a prototype that makes 60 requests in a day hits HTTP 429 and stops, even though the model itself is “free” and the requests are small. The limit is per account, not per model, so spreading requests across multiple free models does not help. The 1,000/day ceiling after $10 is workable for light development, but it is still a request-count meter rather than a token or cost meter, which means a batch of long-context requests and a batch of short ones consume the same quota. Free models are also not guaranteed: the catalog changes frequently, and a model can be removed or rate-limited without notice, which makes them unsuitable as a dependency for anything that needs to run tomorrow.
- In-region routing requires Business (8% fee), not Standard · EU and US in-region routing—sending requests to eu.openrouter.ai or us.openrouter.ai so that prompts and completions are processed entirely within the selected region—is available only on the Business plan, which carries an 8% platform fee on credit purchases instead of 5.5%. Standard users cannot enable in-region routing at any price. This is a structural constraint for European and regulated-industry teams: the feature that most directly addresses data-sovereignty concerns is gated behind a plan that costs 45% more on the platform fee. Enterprise adds contractual SLAs and org-wide policy enforcement on top, but the regional routing itself begins at Business. A team that starts on Standard and later needs EU residency must migrate to Business, accepting the higher fee on every future top-up.
- The gateway is managed, not self-hostable, and uptime is bounded by upstream providers · OpenRouter is a managed service with no self-hosted option. Teams that need data residency beyond EU/US routing, full infrastructure control, or the ability to run the gateway inside their own VPC must choose a different tool—LiteLLM (self-hosted, MIT-licensed) or Portkey (open-source gateway with a hosted tier). OpenRouter's automatic fallback mitigates upstream provider outages by routing to an alternate provider when one fails, but the fallback itself depends on OpenRouter's infrastructure being available. When OpenRouter has an incident, every model behind it becomes unavailable simultaneously, which is a different failure mode from direct provider outages that affect one vendor at a time. The trade-off is deliberate—zero operational burden in exchange for a single point of dependency—but it is a trade-off, not a free lunch, and it matters for teams whose uptime requirements exceed what a managed third-party gateway can contractually guarantee on self-serve tiers.
- The data-improvement opt-in is framed as a 1% discount, which blurs the privacy boundary · OpenRouter's default data posture is genuinely privacy-friendly: prompts and responses are not stored unless the user opts in to private logging or OpenRouter's product-improvement program, or content is flagged for CSAM review. The product-improvement opt-in is off by default and offers a 1% discount on model usage in exchange for allowing OpenRouter to use prompts and completions to improve the product. The distinction that matters—and that the discount framing tends to blur—is between using data to improve OpenRouter's routing, ranking, and observability features, and using it to train models. OpenRouter states it does not use logged data for model training, and that logged data is stored solely for the user's own review; the 1% program is for product improvement, not model training. The boundary is documented, but it requires reading two separate pages to understand, and a user who accepts the 1% discount without reading the fine print is making a data decision on the basis of a pricing incentive rather than a privacy assessment.
Key Features
- 500+ models across 80+ providers through one OpenAI-compatible endpoint
- Automatic fallback when a provider fails, configurable per request or per provider
- BYOK (bring your own provider keys) with prioritized, fallback, and shared-capacity modes
- Typed SDKs in TypeScript, Python, and Go with identical streaming, tools, and structured output behavior
- Agent framework (@openrouter/agent) for multi-turn loops and provider-routed agents
OpenRouter vs. Portkey and LiteLLM
This comparison uses official pricing pages, documentation, and independent gateway comparisons available as of November 2026. It focuses on the structural difference that determines which gateway a team should choose: what each product optimizes for. OpenRouter optimizes for model breadth and zero-ops access; Portkey (now part of Palo Alto Networks) optimizes for guardrails, compliance depth, and enterprise governance; LiteLLM optimizes for self-hosted control and data sovereignty. They overlap on the basics—one API, multi-provider routing, failover—but the choice is usually determined by which control the team cannot compromise on, not by which has the longest feature list.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| OpenRouter Official BYOK documentation (5% fee and $25K allowance) Official data collection policy (default no storage, opt-in logging) Official developer platform (500+ models, MCP server) | A managed LLM gateway optimized for model breadth and zero-ops access, with 500+ models across 80+ providers behind one OpenAI-compatible endpoint and automatic fallback. | The widest model catalog and the lowest operational burden in the category. No self-hosting, no native guardrails beyond ZDR and in-region routing, and no contractual uptime guarantee on self-serve tiers. The fee structure is credit-purchase-based (5.5% card, $0.80 minimum; 8% Business; 5% crypto) rather than per-request, which makes small top-ups disproportionately expensive. BYOK is free to $25,000/month, which is a non-issue for almost all users. In-region routing and contractual SLAs require Business or Enterprise. | Developers and teams that want the broadest managed model access with the least infrastructure work, are comfortable with a managed single point of dependency, and do not need self-hosting or native guardrails beyond ZDR and regional routing. | Free: 25+ free models, 20 req/min, 50 req/day. Standard: token passthrough, 5.5% card fee ($0.80 min), 5% crypto, BYOK free to $25K/month. Business: 8% fee, EU/US in-region routing. Enterprise: custom, SSO/SCIM, SLAs, ZDR guarantees, custom BYOK allowance. |
| Portkey Portkey feature comparison and pricing (official) Independent gateway comparison (documents guardrail and compliance positioning) | An AI gateway now part of Palo Alto Networks' Prisma AIRS, optimized for guardrails, compliance depth, and enterprise governance rather than model breadth or zero-ops simplicity. | Portkey's differentiation is native guardrails (20+ built-in, plus LLM-based and partner checks), RBAC, audit trails, and integration with PANW's AI Runtime Security for real-time prompt and response blocking. The gateway core is open source (MIT) and self-hostable; the hosted platform has a free Dev tier (10k requests/month) and a Pro tier at $49/month. Portkey does not compete with OpenRouter on model catalog breadth—its value is what it does before and after the model call, not how many models it can reach. A team that needs to block risky prompts, enforce organizational policy, and pass an audit will find OpenRouter's ZDR routing and in-region options insufficient; a team that primarily needs cheap access to many models will find Portkey's guardrail focus to be overhead it does not need. | Enterprises in regulated industries that need native guardrails, RBAC, audit trails, and compliance depth at the gateway layer, and that are willing to trade model breadth and zero-ops simplicity for governance features they cannot build themselves. | Open-source gateway: free (MIT), self-hostable. Hosted Dev: free, 10k requests/month. Pro: $49/month with RBAC and service accounts. Enterprise: custom. Overage on paid plans at $9/month increments. |
| LiteLLM LiteLLM pricing (official) Independent comparison (documents self-hosted control and operational costs) | An MIT-licensed, self-hosted LLM gateway and Python SDK that puts 140+ providers behind one OpenAI-compatible API, optimized for data sovereignty and full infrastructure control. | LiteLLM is the self-hosted counterpoint to OpenRouter's managed model. The open-source gateway is free to self-host with unlimited requests, and it can run inside a VPC or on-premises, which is the only option in this comparison that satisfies strict data-residency and network-control requirements without a contractual negotiation. The trade-off is operational: LiteLLM is software you run, not a service you consume, and self-hosting means owning upgrades, scaling, and 24/7 availability. Enterprise features—SSO/SAML, SCIM, RBAC, audit logs—require a paid license (roughly $30k/year or $500–$1,000+/month depending on deployment), and hidden operational costs of $200–$500/month in infrastructure are typical. OpenRouter gives you breadth with zero ops; LiteLLM gives you control with real ops burden. | Teams with SRE capacity that need self-hosted control, VPC or on-premises deployment, or data-residency requirements that a managed gateway cannot contractually satisfy, and that are willing to own the operational burden of running the gateway themselves. | Open-source gateway: free to self-host (MIT), unlimited requests. Enterprise license: roughly $30k/year or $500–$1,000+/month depending on request capacity and deployment architecture. Hidden infrastructure cost: $200–$500/month typical. |
Comparison checked against official product information on 2026-11-04. Features and pricing may change.
Frequently Asked Questions
Does OpenRouter actually charge no markup on tokens?
Yes and no. The per-token price for each model matches the provider's published list rate—that part of the claim holds. The fee is charged when you buy credits: 5.5% by card ($0.80 minimum) on Standard, 5% for crypto, 8% on Business. The $0.80 minimum means a $5 top-up pays 16% effectively, a $10 top-up pays 8%, and the fee only approaches 5.5% around $15+. The token price is unmorked; the cost of loading money is not, and it is disproportionate at small top-up sizes.
Is BYOK really free?
For almost everyone, yes. BYOK is free up to $25,000 per month of list-price inference on the Standard plan, then 5% of the excess. Enterprise defaults to $200,000 per month. Most developers and small teams never approach those thresholds, so the 5% fee is a non-issue. The real cost of BYOK is operational: managing provider keys, understanding prioritized versus fallback routing, and accepting that requests through your keys bypass OpenRouter's rate-limit management. If you are spending more than $25,000/month at list price, the 5% tail applies and should be modeled.
How do free models work on OpenRouter?
Models whose IDs end in :free cost nothing per token and are limited to 20 requests per minute and 50 requests per day. The daily limit rises to 1,000 if the account has purchased at least $10 in credits at any point in its lifetime—not a subscription, just a one-time top-up that remains usable on paid models. The limit is per account, not per model, and it is a request-count meter rather than a token meter. Free models can change or disappear without notice, so they are suitable for experimentation and light development, not as a stable production dependency.
Can I get EU or US data residency on OpenRouter?
Yes, but only on Business or Enterprise. In-region routing sends requests to eu.openrouter.ai or us.openrouter.ai so that prompts and completions are processed entirely within the selected region. Business carries an 8% platform fee on credit purchases instead of 5.5%. Enterprise adds contractual SLAs, SSO/SAML, SCIM, and org-wide policy enforcement. Standard users cannot enable in-region routing at any price; the feature is gated behind the higher-fee plan.
Does OpenRouter train on my prompts?
Not by default. OpenRouter does not store prompts or responses unless you opt in to private input/output logging or to OpenRouter's product-improvement program, or content is flagged for CSAM review. The product-improvement opt-in is off by default and offers a 1% discount on model usage in exchange for allowing OpenRouter to use prompts and completions to improve its product—not to train models. Logged private data is stored for your own review and is not accessed by OpenRouter for training or analytics. CSAM screening can trigger retention regardless of settings. Enterprise adds ZDR routing with contractual guarantees.
Alternative Tools
-
Context7 Context7 is an Upstash-built MCP server that injects up-to-date, version-specific library documentation and code snippets into AI coding agents, reducing deprecated-API hallucinations for indexed frameworks and packages. -
Firecrawl Firecrawl is the most complete LLM-ready web data API: six endpoints covering search, scrape, crawl, map, parse, and interactive browser sessions, with clean Markdown output and first-class MCP support. The advertised 1-credit-per-page rate is a base rate—JSON extraction, PDF parsing, ZDR, and prompt-injection checking multiply it by up to 9x, and Interact bills by the browser minute. Self-hosting is AGPL-3.0 but omits the anti-bot, Agent, and Interact features that define the cloud product. Plan credits do not roll over below $599/month. -
Composio Composio is a managed tool-integration and authentication layer that exposes 1,500+ apps as agent-callable tools via SDKs, a CLI, or hosted MCP endpoints, handling OAuth, tool search, and execution for AI agents. -
Google AI Studio Google AI Studio is the fastest zero-friction way to get a working Gemini API key and prototype multimodal prompts in a browser, with no credit card. The catch is structural: the free tier excludes Gemini 4 Argon and all image/video models, trains on your inputs with no self-serve opt-out, and hides rate limits inside a console; paid rates for Flash are promotional and double in 2027; thinking tokens are billed as output; and AI Studio has no IAM, VPC-SC, SLAs, or regional endpoints—those live in Vertex AI, which is a separate migration. -
Model Context Protocol MCP is the de facto open standard for connecting AI agents to tools and data, adopted by every major AI platform and stewarded by the Linux Foundation. It solves the integration transport problem well. It does not solve the security problem: tool poisoning succeeds at 72.8–84.2% rates against leading models, the specification has no context isolation or least-privilege enforcement, and 49% of enterprises cite security and compliance as their top deployment challenge. -
Exa Exa is a semantic search API built on its own 500B+ page index, with page text included in search results, strong entity-lookup accuracy, an official MCP server, and per-request pricing from $4/1k. It is not a benchmark leader across the board—it is statistically tied with Brave and Firecrawl on independent agentic search benchmarks—and its 10 QPS limit, non-paginating search, crawl-bounded recency, and Enterprise-only compliance features are hard constraints that require workflow design around them.