Anthropic’s lineup has four tiers at any given time: a small fast model (Haiku), a mid model that is very good at coding and agentic work (Sonnet), the default for serious judgment (Opus), and a top tier for the hardest long-running work (Fable). Names and version numbers move; the shape of the choice does not. Check Anthropic’s models page for current IDs and pricing.
The decision rule
- Is it allowed at all? Coursework under an honor code, a client’s data that should not leave their systems: stop here, no model is the right model.
- Does a wrong answer cost real money or reputation? Client tax logic, a resume, statistics that will be published: Opus. The cost difference on low-volume, high-stakes work is pennies.
- Is it above what Opus handles, or a long autonomous run? A whole-codebase refactor with the full spec up front, a multi-hour research task: Fable. Using it for routine edits is paying a lot more for nothing.
- Otherwise, Sonnet. Trivial lookups, classification, boilerplate: Haiku.
How I route my own work
| Work | Tier |
|---|---|
| Pipeline and categorization logic for client financial data | Opus |
| Routine dashboard work: components, CSS, config | Sonnet |
| Bayesian model design and PyMC debugging | Opus (stats subtleties are where weaker models mislead confidently) |
| Sync scripts, environment files, README polish | Sonnet |
| Editing my own writing | Sonnet for line edits, Opus for a real critique of an argument |
| Subagents that only read files and count things | Haiku |
Two things that change the math
- Effort is a second dial. The larger models take an effort setting (low through max). Lower effort means fewer tool calls and shorter output. Try one step lower than you think you need before reaching for a bigger model.
- On a subscription you are spending usage, not dollars. The logic is identical; the top tier just drains your weekly limit much faster.
