Routing : Haiku, Sonnet, Opus, or Fable? A Practical Routing Table for 23 Real Developer Tasks
Most developers pick one Claude model and use it for everything — the fast track to either overspending or underperforming.
Quick Navigation
Difficulty: Intermediate
Estimated Time: 5-10 minutes
Prerequisites: Familiarity with Claude models, Basic software development workflow, Understanding of AI coding assistants, Awareness of reasoning effort settings
Stop paying Fable prices for grep-level work — and stop asking Haiku to design your database.
Most developers pick one Claude model and use it for everything — the fast track to either overspending or underperforming. Anthropic now ships four tiers — Haiku 4.5, Sonnet 4.6, Opus 4.8, and Fable 5 — each with adjustable reasoning effort. The right pairing can slash costs on simple tasks and unlock far better results on hard ones. Over one working session, I routed 23 real development tasks — from "find a variable" to "architect a repo from scratch" — through a simple scoring lens. The table below is the result: every task, sorted from the lightest pairing to the heaviest.
The routing table (sorted by model / effort)
| # | Task | Model / Effort |
|---|---|---|
| 1 | Search which files are impacted on a repo | Haiku 4.5 / low |
| 2 | Start the application | Haiku 4.5 / low |
| 3 | Search all occurrences of a variable | Haiku 4.5 / low |
| 4 | Log a console error, debug, fix, retest | Sonnet 4.6 / medium |
| 5 | Open a website via MCP, test features, fix them | Sonnet 4.6 / medium |
| 6 | Check Vercel deployment with CLI, fix errors | Sonnet 4.6 / medium |
| 7 | Fix orthography/text language across all site pages | Sonnet 4.6 / medium |
| 8 | Design a scoring rubric / decision framework | Sonnet 4.6 / medium |
| 9 | Write a blog article with up-to-date content | Sonnet 4.6 / high (+ web search) |
| 10 | Plan a new feature as a single .md file | Opus 4.8 / high |
| 11 | Develop a feature + integrate in repo, same coding style | Opus 4.8 / high |
| 12 | Upgrade all package dependencies, test, fix impacts | Opus 4.8 / high |
| 13 | New feature impacting many files/features | Opus 4.8 / high (Fable 5 if core architecture) |
| 14 | Web design / UX enhancement of the entire site | Opus 4.8 / high |
| 15 | Security analysis + fix | Opus 4.8 / high (Fable 5 if full audit of a large app) |
| 16 | Research tech components, compare + score in a table | Opus 4.8 / high (+ web search) (Fable 5 if foundational) |
| 17 | Law consultation for technologies | Opus 4.8 / high (+ web search) (Fable 5 if multi-jurisdiction) |
| 18 | Multi-aspect scoring problem | Opus 4.8 / high (Fable 5 if critical decision) |
| 19 | Write tests for 100% coverage of the whole application | Opus 4.8 / high (Sonnet if small app, Fable if huge) |
| 20 | Security audit of a large multi-aspect application | Fable 5 / high |
| 21 | Database schema/model architecture spanning many features | Fable 5 / high (Opus if small/simple app) |
| 22 | Mathematical modeling + optimization + algorithm choice | Fable 5 / high (Opus if textbook problem) |
| 23 | Architect a new repo from scratch for a complex large-scale app | Fable 5 / high |
How to read it (Medium-friendly summary)
Haiku 4.5 / low — mechanical work. Searches, lookups, starting apps, renames. One obvious answer, zero ambiguity. Tasks 1–3.
Sonnet 4.6 / medium — daily dev work. Routine debugging, deployment fixes, text cleanup, well-scoped conceptual tasks. Tasks 4–8.
Sonnet 4.6 / high + web search — fresh-content writing. Anything whose value depends on current information. Task 9.
Opus 4.8 / high — multi-file complexity. Feature planning and development, dependency upgrades, site-wide UX, security fixes, multi-criteria comparisons, full test coverage. Tasks 10–19.
Fable 5 / high — foundational reasoning. Large-scale audits, cross-feature schema design, mathematical optimization, greenfield architecture. Expensive to get wrong, hard to change later. Tasks 20–23.
Key insight: Effort tracks blast radius. The moment a mistake becomes expensive to undo, the task earns "high" — regardless of how short the prompt looks.
Conclusion
The pattern is simple: mechanical work goes to Haiku, daily dev work to Sonnet, multi-file complexity to Opus, and foundational or mathematical problems to Fable. Effort follows the same gradient — low for lookups, medium for routine fixes, high whenever ambiguity or blast radius grows. Add web search whenever the answer depends on a changing world: laws, package versions, prices, news. When in doubt, score the task on five axes — scope, ambiguity, interdependencies, cost of mistake, reasoning depth — and let the total decide. Route deliberately, and the model stops being a setting; it becomes part of your engineering judgment.