OpenRouter: one API, many models
OpenRouter: Model Routing, Pricing and What the Rankings Show
How OpenRouter gives one API across hundreds of models, how its fees and routing work, what its public usage rankings say about which models teams actually run, and where a gateway fits in an AI stack.
Prices and versions: figures are published list prices in US dollars per 1 million tokens, read from each provider’s own pages on 18 November 2026. Models, prices and promotions change often, so confirm on the provider’s page before you budget.
What OpenRouter is
OpenRouter is a gateway: one API and one bill in front of hundreds of models from OpenAI, Anthropic, Google, open-weight labs and many hosting providers. You call a model by an id of the form provider/model-slug, such as a dated slug for a specific release, and OpenRouter sends the request to a provider that serves it. If a provider returns an error, it falls back to the next provider automatically.
Fees and plans
OpenRouter states that it adds no markup to inference prices; it earns through a fee on credit purchases (5.5 percent with an $0.80 minimum, or 5 percent for crypto). Its pricing page lists the plans below. The FAQ we read mentions only the 5.5 percent figure, so confirm the plan fee that applies to you.
| Plan | Platform fee on credit purchases | Notes |
|---|---|---|
| Free | Not applicable | 25+ free models, 4 free providers, low daily request limits |
| Standard | 5.5% | 600+ models, 80+ providers, auto-routing and preferred vendor selection |
| Business | 8% | Adds in-region EU and US routing |
| Enterprise | Discounts available | Custom limits, invoicing, SSO, SLAs, dedicated support |
Bringing your own provider keys is covered up to $25,000 a month of list-price inference with no fee, then a 5 percent fee. Credits expire after a year, and refunds are limited to the first 24 hours.
Routing options and variants
- Provider fallback. Automatic failover to the next provider when one errors.
:nitroPrefer the fastest throughput.:floorPrefer the lowest price.:exactoPrefer providers with more reliable tool calling.:freeand:batchFree-tier variants with their own rate limits, and batch pricing.- Free router. A free router picks a free model for you. Free usage is limited to 50 requests a day, or 1,000 a day once you have bought at least 10 credits.
- Web search. The old online variant is deprecated in favour of a web-search server tool.
What the usage rankings show
OpenRouter publishes weekly token usage by model. In the week to 16 November 2026 the top ten looked like this. Usage measures volume, not quality, and it reflects the developers and agent tools that use OpenRouter rather than the whole market.
| Rank | Model | Lab | Weekly tokens | Change |
|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | DeepSeek | 61.2T | +10% |
| 2 | Hy4 | Tencent | 16T | 0% |
| 3 | DeepSeek V4.1 Flash | DeepSeek | 15.5T | -12% |
| 4 | GLM 5.3 Flash | Z.ai | 11.8T | -3% |
| 5 | GPT-6 Luna | OpenAI | 8.08T | +14% |
| 6 | Claude Opus 5.5 | Anthropic | 7.37T | +9% |
| 7 | Step 5 Preview | StepFun | 6.36T | -11% |
| 8 | MiMo-V2.6-Flash | Xiaomi | 4.16T | +63% |
| 9 | GLM 5.3 | Z.ai | 3.74T | +5% |
| 10 | GPT-6.1 Sol | OpenAI | 3.71T | -5% |
Seven of the ten are low-priced models from Chinese labs, and the three from US labs include GPT-6 Luna, the cheapest tier in OpenAI’s line. The pattern is consistent with cost-sensitive, high-volume agent traffic, where the lowest price per solved task wins. On OpenRouter’s own models page, we also saw Tencent Hy4 flash listed at $0.035 input and $0.1405 output per million tokens, GPT-6.2 Sol at $2 and $10, and Qwen4 Max at $1.425 and $4.275.
When a gateway makes sense
- Many models, one integration. Swap models without rewriting clients, and test cheaper models on live traffic.
- Resilience. Provider fallback and multiple hosts reduce single-provider outages.
- Cost control. Route simple tasks to cheap models and escalate hard ones.
- Trade-offs. A third party sits in the data path, credit fees apply, and some providers log prompts. Check data-handling settings, regional routing and your own data and export rules before sending regulated or sensitive data.
Where this meets infrastructure
For production we often put a gateway or router inside the customer’s own boundary, mixing commercial APIs (OpenAI, Anthropic), Hugging Face endpoints and self-hosted open models on hosted or neocloud capacity. The agent platform and security controls sit on top.
Where this fits in the stack
Common questions
Does OpenRouter mark up model prices?
OpenRouter says it adds no markup to inference prices and instead charges a fee when you buy credits (5.5 percent with an $0.80 minimum on its FAQ, with other plan fees on its pricing page).
What are the :nitro, :floor and :free variants?
They are suffixes on a model id: :nitro prefers speed, :floor prefers the lowest price, :free uses free-tier access with its own limits, and :exacto prefers reliable tool calling.
Do OpenRouter rankings show the best models?
No. They show token volume. Popular models are often the cheapest or the default in popular agent tools, so use the rankings for adoption signals and run your own evaluation for quality.
Is OpenRouter suitable for sensitive data?
It depends on your data rules. Providers differ in logging and retention, so review the privacy settings, regional routing options and your compliance obligations before sending regulated data.
Sources and dating
Source: OpenRouter models, rankings, pricing and FAQ pages, read 18 November 2026. Rankings are weekly token counts to 16 November 2026. Per-model prices change constantly on OpenRouter, so we show only a few examples. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against the provider pages before we size or budget a build. OpenRouter and the model names are trademarks of their owners. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: LLM models and token costs · OpenAI models · Anthropic Claude models · AI application layer
Want a model gateway inside your own boundary?
Model routing design
