Anthropic Releases Model Selection Guide: Haiku 5.5 Saves Tokens on Routine Work

Everyone is used to thinking that the more powerful the model, the better the result. In practice, this is a quick way to burn through your limits: a heavy model reads logs and searches for files at the same cost as when it designs architecture. Anthropic has released a guide that tackles this problem and suggests which model to use for which job.

Which model for what

The guide splits the lineup into four roles. Opus 5.5 is the all-round option, and Anthropic advises starting most tasks with it. Haiku 5.5 is the fastest and cheapest: classification, simple operations, and subagents that read logs. Sonnet 5.5 suits everyday work: bug fixes, documents, conversation. Fable 5.2 is the most powerful model in the lineup, best reserved for large-scale projects, complex research and scientific calculations.

The main idea is simple: don't pay for power where it isn't needed. The savings are noticeable, while quality barely drops.

The Explore subagent trap

The most practical detail concerns Claude Code. The built-in Explore subagent, which searches the codebase, used to run on the cheap Haiku. Now it uses the current session's model. If you are on Fable 5.2, every repository search is billed at its rate, and your limits vanish unnoticed.

The fix is to create your own subagents. Anthropic even offers a ready-made prompt that Claude will execute itself. It first shows the Claude Code version, the provider and the contents of the agent folders, and overwrites nothing without your "yes". It then creates three agents with model: haiku, with read-only and command-execution permissions:

• Explore returns path:line and a short explanation instead of whole files;
• log-reader parses build logs and names the first cause of failure with the exact error text;
• test-runner runs tests and replies "passed / failed" with a list of failures.

You don't need to set the CLAUDE_CODE_SUBAGENT_MODEL variable: it would also make Plan and general-purpose cheaper, and those are exactly the ones that need a strong model.

Pitfalls

The haiku alias means Haiku 5.5 only on the Anthropic API with Claude Code v2.1.293 and newer. On Bedrock, Google Cloud, Foundry and Claude Platform on AWS, it still points to Haiku 4.5. There you need to set ANTHROPIC_DEFAULT_HAIKU_MODEL manually, if the new version is available from the provider.

Another rule from the guide: delegate only what is expensive in volume but simple in meaning. Tasks of one to three tool calls are better done yourself. Edits, decisions and reviews stay with the main model, and the subagent's output should be checked at the specified location before editing.

What to do in practice

The escalation order is as follows: if a task fails, first verify the result with tests, then raise effort one level, and only then switch to a stronger model. By default, Anthropic recommends Opus 5.5 with effort medium, and high for difficult tasks. Sonnet 5.5 is good for work with a clear specification and for documents. Fable 5.2 is for large migrations, architecture and long autonomous runs. Before switching models, it helps to run /compact so you don't carry over excess context.

For teams that pay for tokens, this is one of the cheapest ways to cut the bill: half an hour to set up three subagents and rules in CLAUDE.md. Check the result in /agents: Explore should show Haiku. If not, you are still paying flagship prices for search.