32.3 Choosing small vs large models strategically

When to Use Each

Model Cascade

// Try cheap model first, escalate if needed
async function cascadeGenerate(prompt: string) {
  // Step 1: Try Flash
  const flashResponse = await flash.generate(prompt);
  
  // Step 2: Check confidence
  if (flashResponse.confidence > 0.9) {
    return flashResponse;  // Flash is good enough
  }
  
  // Step 3: Escalate to Pro
  console.log('Escalating to Pro due to low confidence');
  return pro.generate(prompt);
}

// Result: 80% of requests handled by Flash (cheap)
// Only 20% need Pro (expensive but necessary)

Smart Routing

// Route based on task complexity
function selectModel(task: Task): string {
  const complexity = estimateComplexity(task);
  
  if (complexity === 'simple') {
    return 'gemini-1.5-flash';  // $0.075/1M
  } else if (complexity === 'medium') {
    return 'gemini-1.5-flash';  // Still use Flash
  } else {
    return 'gemini-1.5-pro';    // $1.25/1M
  }
}

function estimateComplexity(task: Task): string {
  if (task.type === 'classification') return 'simple';
  if (task.type === 'extraction') return 'simple';
  if (task.requiresReasoning) return 'complex';
  if (task.outputLength > 1000) return 'complex';
  return 'medium';
}

Where to go next

• 32.4 Batch processing vs interactive