Prompt Classifier

How it works

The classifier analyzes your prompt using static heuristics (zero latency) and assigns a complexity level. Each level maps to the cheapest model that can handle it well.

Signals

The classifier extracts these signals from your prompt:

• Token count — longer prompts tend to be more complex
• Code blocks — presence and line count
• File references — multi-file operations score higher
• Task keywords — refactor/architect vs fix/explain
• Constraints — "must", "ensure", "backward compatible"
• Conversation depth — deep sessions stay on current model

Standalone Package

npm install @costlens/classifier

import { classify } from '@costlens/classifier';

const result = classify({
  prompt: 'Fix the typo in line 3',
  currentModel: 'gpt-5.4',
  provider: 'openai',
  messageCount: 3,
  routingMode: 'balanced',
});
// => { level: 'simple', suggestedModel: 'gpt-5.4-mini', confidence: 0.78 }

Safety Rules

The classifier will never downgrade when:

• It's the first message in a session
• Prompt contains "careful", "production", "critical"
• Conversation is deeper than 15 turns
• The suggested model would be more expensive