OpenRouter is a single API in front of almost every AI model. Instead of opening accounts with Anthropic, OpenAI, Google, Mistral, DeepSeek and the rest, you get 1 key, 1 bill and 1 endpoint. Switching from Claude to Gemini is a change of 1 word in the request.
It's the model equivalent of what treg does for agent tools (see treg - OpenRouter for agent tools ).
26 Sep 2026 from the live models API:
| Models | 458, from 63 labs. OpenAI (100), Qwen (54), Google (41) and Anthropic (27) have the most |
| Providers | 110 hosting companies that serve those models |
| Free models | 17, capped at 50 requests a day |
| Formats | Text, images, video, speech-to-text and text-to-speech |
| API | Compatible with the OpenAI SDK |
For business people
The problem: every AI lab has its own account, key, billing, rate limits and outages. Testing 5 models means 5 contracts. And when your only provider goes down, your product goes down with it.
What OpenRouter gives you:
- 1 key, 1 invoice for hundreds of models.
- Model choice in code, not in contracts. Compare models on the same prompt, and switch without new sign-ups.
- Automatic fallbacks. If a provider has an outage or hits a rate limit, the request goes to another provider of the same model.
- Cost tracking across all providers in 1 place.
- Public usage rankings that show which models developers actually use, a useful market signal.
What it costs
Model prices are the providers' list prices. OpenRouter earns a fee when you buy credits.
| Plan | Fee on credit purchases | Bring your own provider keys |
|---|---|---|
| Standard | 5.5% | Free up to $25,000 of usage a month, then 5% |
| Business | 8% | Same as Standard |
| Enterprise | Discounts on request | Free up to $200,000 a month, then 5% |
Examples of list prices on 26th September 2026, per million tokens:
| Model | Input | Output | Context |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $20.00 | 1M tokens |
| Claude Sonnet 5 | $2.00 | $10.00 | 1M tokens |
| Gemini 3.8 Flash | $0.75 | $3.75 | 1M tokens |
Limits and risks
- A middleman. Your prompts pass through OpenRouter and then a provider. The account settings can block providers that train on your data, and a request can demand zero-data-retention endpoints only. Keeping data inside the EU is an Enterprise option.
- 5.5% on everything adds up at volume. Past a certain spend, a direct contract with 1 lab is cheaper.
- Provider quality varies. The same open model can run slower, or with a different quantisation, at 1 host than at another. Routing hides that unless you pin providers.
When to choose it: prototypes, model comparisons, apps that need several labs, and anything that must survive a provider outage. For heavy use of 1 model, go direct.
For technical people
Quick start
The API mirrors OpenAI's, so the OpenAI SDK works with a different base URL:
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="<OPENROUTER_API_KEY>",
)
reply = client.chat.completions.create(
model="anthropic/claude-sonnet-5",
messages=[{"role": "user", "content": "Summarise this contract in 3 bullets."}],
)
print(reply.choices[0].message.content)
Models are named lab/model. The full catalogue, with prices and context sizes, is public at https://openrouter.ai/api/v1/models, no key needed.
Routing controls
By default, OpenRouter spreads requests across the providers of a model, favouring low prices and skipping providers with recent outages. A provider object in the request overrides that:
| Parameter | Effect |
|---|---|
order |
Try these providers first, in this order |
only / ignore |
Allow or exclude specific providers |
allow_fallbacks |
Set to false to fail instead of switching provider |
sort |
Rank by price, throughput or latency instead of load balancing |
data_collection |
Set to "deny" to skip providers that store your data |
zdr |
Set to true for zero-data-retention endpoints only |
max_price |
Cap the price per token |
2 shortcuts on the model name: :nitro for the fastest providers, and :floor for the cheapest. A models array instead of model gives a fallback chain across different models, for example Claude first, then Gemini.
reply = client.chat.completions.create(
model="anthropic/claude-sonnet-5",
messages=[{"role": "user", "content": "..."}],
extra_body={"provider": {"zdr": True, "data_collection": "deny"}},
)
⚠️ WARNING: the free models are rate-limited to 50 requests a day, and some providers behind free endpoints may log or train on prompts. The account settings control this separately for free and paid models. Don't send anything private to a
:freemodel.
Value
OpenRouter is the fastest way to try a new model the day it launches, and the simplest way to add a fallback to anything in production. For my own setup, most work runs through Claude Code on a subscription, and the local model runs on my own machine (see Fine-tuning your own model ), so OpenRouter isn't a daily tool. It earns its place for side-by-side model tests and for scripts that need a second lab as backup, with zdr switched on for anything that touches client material.
Further Reading