Model Routing Needs Clear Escalation Rules

Model routing becomes an operating concern as soon as a model family offers several price, speed, and capability tiers. OpenAI explicitly recommends matching the model, reasoning effort, and speed to each workload in its GPT-6 guide. Microsoft provides a Foundry router that selects a model for each prompt.

A low-cost default is not enough

OpenAI positions GPT-6.1 Sol between GPT-6 Luna and GPT-6 Astra. According to the provider, API pricing is $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. OpenAI still recommends Astra for the most difficult scientific tasks.

This does not mean that every team should choose models manually. Companies need a default path for frequent tasks and defined reasons for escalation. Microsoft describes Model Router as a separate model that selects a suitable LLM from an approved set in real time. The decision becomes automatable, but not automatically compliant with governance requirements.

Routing needs measurable rules

A robust routing contract should define at least five points:

• Default: Which model and reasoning effort apply to routine cases?
• Escalation: Which task class, failed validation, or protection level triggers a stronger model?
• Budget: Which limits apply to latency and cost per successful task?
• Approval: Which data classes, providers, regions, and model versions may the router use at all?
• Fallback: What happens after a timeout, quality failure, or unavailable target model?

OpenAI recommends testing representative tasks before production and measuring task success, latency, and cost per successful task. These metrics should then become part of routing telemetry. Otherwise, the router may optimize token prices while rework or incorrect decisions increase.

Additional checks for DACH companies

If a router includes models from several providers, privacy, contractual basis, and the intended processing region must be assessed for every permitted path. Technically effective routing must not open an unapproved region or data class.

Start with a small number of task classes and a fixed escalation matrix. Model routing is not an invisible optimizer. It is an operating decision that must combine quality, cost, and risk per task in a traceable way.