One stable alias fronts a private model mixture.
The public API model ID is broken-one and the product name is BrokenGPT. Requests are privately routed across a mixture of open-source models. Backend checkpoint and provider identities can change and are deliberately excluded from public responses and metadata.
Current service-level specifications.
| Field | Current public contract | Note |
|---|---|---|
| Model alias | broken-one | Stable public identifier |
| Model type | Privately routed mixture of open-source models | Checkpoint identities are private |
| Context window | 1,048,576 tokens by default | Live deployment configuration is shown on the system page |
| Maximum output | Caller-selectable within remaining context | No separate fixed public maximum is advertised |
| Streaming | Supported | SSE on chat completions and the supported Responses subset |
| JSON output | Supported, non-streamed | json_object and JSON Schema validation |
| Function tools | Supported | Client executes returned tool calls |
| Image input | Not advertised | Current public message content is text |
| Endpoints | Chat completions, Responses, batches, token count, usage, models | Each has a documented subset |
| Knowledge limit | Not published as one fixed date | Routing can change; verify current facts |
| Rate limits | Plan and key dependent | Returned in headers and account controls |
| Pricing | Input and output token based | See the live pricing page |
Behavior must be evaluated separately from the schema.
The gateway does not add a keyword censorship filter or platform-authored refusal responder. Every valid prompt reaches configured inference, but the model mixture can still produce refusals, errors, hallucinations, incomplete answers, or insecure code.
- Use a representative evaluation set for your domain.
- Measure correctness and refusal as separate axes.
- Record serving date, prompt, parameters, latency, and cost.
- Re-run regression prompts after meaningful product changes.
The API contract is OpenAI-compatible, not provider-equivalent.
Applications using the supported fields may only need to change the base URL, key, and model alias. Other endpoints, multimodal inputs, hosted tools, and exact output behavior should not be assumed. Consult the compatibility matrix and live documentation before deployment.