How ANT Routing Works

The request is analyzed to determine its task type, complexity, and requirements. This happens in milliseconds using a combination of heuristics and a custom classifier LLM.

• •Heuristic pass detects vision content, context window needs, and reasoning indicators
• •Custom classifier LLM assigns task family tags (code, creative, reasoning, analysis, etc.)
• •Stage 2 LLM escalation for ambiguous requests below the confidence threshold
• •Fallback to heuristic classification if the classifier is unavailable

Every eligible model is scored against the classified request. Scores combine historical performance data, task-specific success rates, latency, cost, and provider health.

• •Intelligence score: weighted composite of success rate, latency, cost efficiency, routing popularity, and task versatility
• •Task affinity: models get bonus scores for task families they excel at
• •Provider health: circuit breaker state, recent error rates, and fallback history
• •Cost awareness: free models get a boost when quality is comparable to paid alternatives

The highest-scoring model is selected, with fallback candidates ranked for automatic retry. The circuit breaker can skip unhealthy providers entirely.

• •Top-scored model is selected as primary, next best as fallback
• •Circuit breaker skips providers with recent consecutive failures
• •Reasoning requests route to thinking-capable models with reasoning_effort wiring
• •Direct routing bypass for providers like DeepSeek and NetMind when explicitly requested