Anthropic Publishes a Claude Model Selection Guide: Start Most Coding on Opus 5.5 at Medium, Tune Effort Before Switching Models, Put Subagents on Haiku 5.5

Core Background and Industry Pain Points

After Fable 5.2 and Mythos 5.2 launched, the Claude lineup has four generally available models, Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.2, with input prices ranging 80x from $0.10 to $8 per million tokens. Many developers are unsure which to default to, so Anthropic published an official selection guide by Addy Osmani on claude.dev; the ClaudeDevs post passed 12,000 bookmarks.

Architecture Highlights and Underlying Mechanisms

The guide treats choice as two dials: model and effort. Start on Opus 5.5 at medium and raise to high if needed; beyond high, one more effort level can cost as much as a bigger model, so test the next model up at a lower effort instead. The prompt cache belongs to the model that wrote it, so the first turn after a switch pays cache-write again; run /compact first. Thinking blocks are readable only by the same or higher tier, so moving down loses earlier reasoning.

Authoritative Benchmarks and Measured Results

All figures are Anthropic's own runs (vendor figures): on FrontierCode 2.0, Fable 5.2 at high scores 45.3% vs Opus 5.5 at max 45.5% at about half the cost per task; on Terminal-Bench 3.0, Fable 5.1 at max scores 58.0% vs Opus 5.5 at high 58.9% with half the tokens; on SWE-bench Pro, Opus 5.5 at low with failures re-run at high passes about 97% at about $0.17 per task. No third party has reproduced these yet.

Developer Hands-On Guide

In Claude Code, set model: haiku on subagents or set CLAUDE_CODE_SUBAGENT_MODEL; the built-in Explore subagent now follows the session model, so define your own Explore subagent to keep it cheap. Before moving to Fable 5.2, keep history append-only, send tool_choice auto with strict tools, and read content blocks by type. Then compare score against spend on your own tasks at two or three effort levels before committing.