Route Cheap, Run All Day

Heavy Claude Code users hit limits and watch costs climb for the same reason: every task gets sent to the most powerful model, whether it needs it or not. The obvious fix - just pay for more - scales your cost linearly with usage, so the more you automate the more you pay. The smarter fix is to route. Send the hard stuff to Sonnet or Opus and everything else to Haiku, which is fast, cheap, and handles the bulk of routine work without complaint.

You need an Anthropic API key from the Console. Once you have one, this quick test confirms Haiku is reachable before you build anything on top of it.

bashcurl https://api.anthropic.com/v1/messages \
  -H "x-api-key: YOUR_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-haiku-4-5-20251001",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "Reply with: Haiku is live."}]
  }'

This is the part people skip - and then wonder why nothing changed. Without a rule, every call still goes to your default model. The logic is simple: reasoning about architecture, debugging a hard error, or writing logic you have not defined yet goes to Sonnet or Opus; reading a file, writing a comment, reformatting, summarising a log, or checking syntax goes to Haiku. Have Claude generate the function for you.

I am building a Claude Code setup with a cheap fallback model (Haiku) to keep costs down and avoid brushing against limits. Write me a Python function called route_model(task_description) that takes a plain-English description of a task and returns either "claude-sonnet-4-6" or "claude-haiku-4-5-20251001". Use these rules: route to Haiku for file reads, summarization, reformatting, syntax checks, boilerplate, and status messages. Route to Sonnet for architecture decisions, complex debugging, multi-step reasoning, and anything involving business logic I have not defined yet. Include a catch-all that defaults to Haiku to minimize cost. Add a comment explaining each routing decision.

Before any agent runs in production, set a hard monthly usage limit in the Anthropic Console - a runaway loop should hit a ceiling, not your card. Set it at roughly three times what you expect to spend so you have room to grow but a worst case that cannot hurt. Then add a lightweight tracker so you can see where your money actually goes.

Write me a Python function called track_api_cost(model, input_tokens, output_tokens) that calculates the estimated cost of an Anthropic API call based on current public pricing for claude-haiku-4-5-20251001 and claude-sonnet-4-6. Store a running total in a local JSON file called api_costs.json with keys: date, model, input_tokens, output_tokens, estimated_cost_usd, and cumulative_monthly_cost_usd. If the cumulative monthly cost exceeds a MONTHLY_CAP variable I can set at the top of the file, print a warning and return False to halt the call. Otherwise return True. Keep the code under 60 lines.

• No routing rule, so everything still goes to your big model - you hit limits just as fast and costs stay high. Build the rule in step 2 before running anything.
• No spend cap set, so a bad loop can run up a large bill overnight. Set the cap before your first real run, not after.
• Routing genuine reasoning to Haiku - it will answer hard tasks, just often with the wrong answer delivered confidently. Be honest about what needs real reasoning and send those to Sonnet or Opus every time.

Run your normal workflow for seven days, then look at where the spend landed. If Haiku keeps handling tasks that need a Sonnet follow-up, promote those. If Sonnet is doing work Haiku could handle, demote those. One round of tuning typically trims another 20 to 30 percent off an already-reduced cost - and from there the bottleneck stops being cost and starts being what you decide to automate next.