Custo de agente: $/action, prompt cache, cascade routing

import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic();

const response = await client.messages.create({
  model: 'claude-sonnet-4-5',
  max_tokens: 1024,
  system: [
    {
      type: 'text',
      text: '<long system prompt with tool definitions + few-shot examples>',
      cache_control: { type: 'ephemeral' },  // ← marca o cache point
    },
  ],
  messages: [
    { role: 'user', content: 'Última pergunta do usuário (varia)' },
  ],
});

// usage.cache_creation_input_tokens — primeira chamada (custou 1.25x)
// usage.cache_read_input_tokens     — chamadas subsequentes (custaram 0.1x)

async function cascadeAnswer(query: string) {
  // Tier 1: rápido e barato
  const haiku = await call('claude-haiku-4-5', query);
  if (haiku.confidence > 0.85 && validateOutput(haiku)) return haiku;

  // Tier 2: maior, ainda barato
  const sonnet = await call('claude-sonnet-4-6', query);
  if (sonnet.confidence > 0.7 || requiresSimpleReasoning(query)) return sonnet;

  // Tier 3: pesado, só quando precisa
  return await call('claude-opus-4-7', query);
}

// Middleware que mata loops com gasto desconfortável
const BUDGET_USD_PER_TASK = 0.50;
let taskCost = 0;

for (const step of agentSteps()) {
  const result = await runStep(step);
  taskCost += estimateCost(result.usage);

  if (taskCost > BUDGET_USD_PER_TASK) {
    log.warn({ taskId, taskCost }, 'budget exceeded, halting agent');
    return { status: 'budget_exceeded', partial: result };
  }
}