BrokenGPT is one product with two ways to use it.
BrokenGPT is an India-built AI service for direct chat and server-side applications. People can use the web workspace, while developers can send the documented chat-completions request shape through an API.
Both surfaces use the stable BrokenGPT identity and the public API alias broken-one. Behind that contract, requests are privately routed across a mixture of open-source models. Backend checkpoint and provider identities are not part of the public interface.
Lower-refusal describes the gateway behavior.
BrokenGPT does not add a platform-authored keyword blocklist or a separate gateway refusal responder. Routing selects capability and availability; every valid prompt is sent to configured inference.
That does not promise that every answer will be correct, complete, or phrased exactly as requested. Model output can still be inaccurate, biased, outdated, insecure, or naturally unwilling to complete a task. Important results require verification.
Use chat for direct work and the API for products.
- Use web chat for interactive writing, analysis, coding, and research workflows.
- Use the API when your own server, agent, application, or automation needs model output.
- Use the OpenAI-compatible surface when an existing client can set a custom base URL, key, and model alias.
- Use the system page when you need the current context cap, data flow, and public model contract.
Know the product boundary before relying on it.
Compatibility covers documented BrokenGPT endpoints and fields; it is not a claim of identical models or complete platform parity with another provider. The current public route supports text chat, streaming, function tools, structured output, Responses, batches, token counting, and usage reporting within its documented subsets.
Model behavior can change as the private mixture evolves. Keep regression prompts, validate structured output, enforce application-level permissions, and retain human review wherever a wrong answer could cause material harm.