Choose a model
Choose an UnderSky model by routing behavior, response quality, latency, context needs, and token price, then validate the choice with a small request.
Model choice is a product decision. Start with the smallest model that meets your quality target, then move up only when an evaluation shows a gap.
Start with automatic routing
Use auto when you want UnderSky to select an available route for a general request. It is useful for prototypes and fallback paths.
{"model":"auto","messages":[{"role":"user","content":"Summarize this release note."}]}Pick a family
The model directory shows current IDs, prices, and supported endpoints.
- OpenAI models cover general chat and Responses-compatible applications.
- Claude models are useful when your prompts depend on long-form reasoning or Anthropic message semantics.
- DeepSeek models offer lower token prices for many text workloads.
- Use the Models API reference when you need the machine-readable list.
Compare the four constraints
Evaluate each candidate on the same prompt set:
- Quality. Check factuality, instruction following, structured output, and tool-call correctness.
- Latency. Measure time to first token and total completion time under your normal concurrency.
- Context. Keep prompts within the model's available context and leave space for the output.
- Cost. Multiply input and output tokens by the prices on the model page. Include retries in the estimate.
Store the winning model ID in configuration so a catalog update does not silently change a production workflow.
Use endpoint semantics
Use /v1/chat/completions for existing OpenAI chat clients. Use /v1/responses when your application already uses the Responses format. Use /v1/messages for the Anthropic-native request shape supported by Claude and DeepSeek V4.
Read the matching API references before switching endpoints. The model page links every endpoint that accepts that model.
Validate a candidate
Send a small request, record the returned model and usage, and compare it with your evaluation result. Review pricing and status before raising concurrency.
For a complete setup, follow the Quick Start, then add the rate-limit recovery behavior before production traffic.
Migrate an OpenAI integration
Move an existing OpenAI SDK or HTTP integration to UnderSky by changing the base URL, API key, model ID, and operational checks.
Choose a media API workflow
Map image generation, text to speech, audio transcription, and social video extraction to the UnderSky endpoint, billing model, and recovery path you need.