Model selection is project-scoped. The dashboard exposes vendor activation and model-family settings without placing upstream credentials in client applications.
#Project settings
Open a project and choose Models. Changes apply to that project's client keys. A revision check prevents an older dashboard view from overwriting newer routing state.
#Vendor families
| Vendor | Examples |
|---|---|
| Gemini | Gemini 3.x, Imagen, Veo |
| OpenAI | GPT-5.6/5.5/5.4, GPT-4.1/4o, o3/o4, GPT Image 2/1.x |
| Anthropic | Claude Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5, and prior models |
| DeepSeek | DeepSeek V4 Flash and V4 Pro |
| Fish Audio | S2.1 Pro, S2 Pro, and S1 text-to-speech |
| Open-source | Qwen, Llama, Kimi, Mistral, Gemma, and more |
#Provider selection
Hosted platform routing and custom open-source supply have separate rollout controls. A request must match a priced canonical model and an eligible upstream. Provider health and agreement status are checked before dispatch.
#Failure behavior
- Unknown or unpriced models fail before upstream dispatch.
- Usage-limit and balance gates run before the selected upstream call.
- Open-source provider capacity failures do not create seller earnings.
- HALO never substitutes a different canonical model silently.