How Pareto works

Pareto serves GLM 5.3 Flash, DeepSeek V4.1 Flash and MiMo V2.6 Flash on its own GPUs. You pay for the tokens you use.

​

Connect your tools

Add your Pareto endpoint and API key to a router that supports custom OpenAI-compatible providers. Your router sends requests for these models to Pareto. See the integration guides. You can also call the API directly.

​

Served by Pareto

Pareto processes these requests on its own GPUs. When the account has no credits for a request, too many requests arrive at the same time, or the model is at capacity, the API returns HTTP 429. A full model returns model_capacity with a Retry-After header at once, so your router can use another provider. When a temporary failure occurs, such as a GPU outage, the API returns HTTP 503. When a model request fails, the API returns HTTP 502. Your router can retry or use another provider if you configure it to do so. See Troubleshooting for retry guidance.

​

Serving system details

We do not disclose the technical details of our GLM 5.3 Flash serving system. This helps prevent competitors from reverse-engineering our stack.

​

Pay per token

Add a card and buy prepaid credits. Input, cached input, and output tokens use your balance at the published rates.

​

Data handling

Pareto offers zero data retention (ZDR): we do not keep request content after a request finishes. We do not use this content to train or evaluate AI models. We do not sell your data to anyone. See the Zero Data Retention Policy and Privacy Policy for what Pareto keeps and how we use it. If you use a router, its data policy also applies to the requests it handles.