Route each task to the cheapest model that passes your own tests, and escalate only when it fails. On Anthropic's public price list for 19 October 2026, Haiku 5.5 input costs 40 times less than Opus 5.5 input, so using the largest model for everything is usually the most expensive habit.
The guidance from the docs
Anthropic's pricing documentation, in its cost strategies for agents, says to choose Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning. Its Claude Code documentation says Sonnet handles most coding tasks well and costs less than Opus, and suggests specifying a smaller model for simple subagent work.
On the same page, Haiku 5.5 is listed at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, Sonnet 5.5 at $2 and $10, and Opus 5.5 at $4 and $20. Google and OpenAI also list small, mid and large models at different prices. See the price snapshot.
A method that fits in a week
- List your task types. Examples: classify a ticket, extract fields, summarize a call, write a function, plan a refactor.
- Build 20 to 50 real examples per task with the answer you would accept.
- Run each task on the small model first, then mid, then large.
- Score the outputs against your acceptance rule, with a human spot check.
- Pick the cheapest tier that clears your bar for each task.
- Add a fallback: if a cheap model fails a check, retry on the next tier.
Price per token is not price per task
A cheaper model that needs retries, or writes longer answers, can cost more per finished task. A larger model may take fewer steps. Measure total tokens and total cost per accepted result, not list price.
What usually fits where
These are starting hypotheses, not rules. Test them.
- Small: routing, tagging, extraction, short rewrites, simple subagents.
- Mid: most coding and writing, summaries with nuance, tool use in agents.
- Large: ambiguous design decisions, hard debugging, planning across many files.
Route in code, not by habit
Put the model name in config per task type. Log the model with every call. Review cost per task monthly, so a price change or a new model gets noticed. This also protects you if a price moves. See scheduled price changes.
Watch for hidden multipliers
Thinking tokens are billed as output, so a large model on a hard task can generate many tokens before it answers. See thinking tokens.
Quick check
Should I always start with the smallest model?
For tasks with clear right answers, yes, test it first. For open ended work, the saving may not be worth the rework.
How do I know the cheap model is good enough?
Run it on your own examples with an acceptance rule written beforehand. Public benchmarks do not know your data.
Sources
- Anthropic pricing documentation
- Claude Code cost management documentation
- OpenAI API pricing
- Google Gemini API pricing
Last reviewed: 19 October 2026.
