Claude Haiku 4.5

The smallest and cheapest slot in Anthropic's 4.5 lineup shipped on October 15, 2025. Replacing Sonnet or Opus was never the point. Haiku 4.5 was built to take over the slice of Claude Code work that never needed either of them. And Claude Code learned to route that work on its own.

Key Specs

The rough split is this. About 30-40% of Claude Code tasks are light: scanning files, short edits, boilerplate, and one-liners. Haiku 4.5 was sized for that slice, and sized to run fast and cheap.

Smart model switching. This is the headline. Rather than picking a model yourself, Claude Code started sending the light work to Haiku 4.5 on its own: quick edits and file reads, obvious one-liner questions, routine boilerplate, the easy refactors. Routing happened under the hood. You kept working while the choice was handled for you.

Pro plan availability. Claude Code Pro users got Haiku 4.5 too, which put automatic routing in reach without the API bill. Pro picked up a real bump in practical value.

Speed. The Haiku family is fast by design. When you are tapping your foot waiting on a one-line answer or a small edit, the latency gap shows up right away. On Sonnet, those replies took 3-5 seconds. Haiku cleared them in under 2 seconds.

Cost reduction. Teams on the API pay per token. Sending the routine work to Haiku knocked daily bills down hard. Scale that across a whole team running Claude Code and the savings pile up fast.

Routine development work is where Haiku 4.5 lands cleanly:

• File reading and summarization. Scan a file, then answer questions about what is inside.
• Quick edits. Small, precise changes in code that already exists.
• Simple questions. "What does this function do?" or "What's the correct import path?"
• Boilerplate generation. Standard patterns, test scaffolds, and config files.
• Commit messages and documentation. Clean commit messages in the conventional format, plus inline doc comments.

Know where it breaks down. Do not hand Haiku:

• Complex architecture decisions. Multi-system design needs the deeper reasoning of Sonnet 4.5 or Opus.
• Multi-file refactoring. When an edit sweeps many files at once, it needs more context held in working memory than Haiku can carry in one pass.
• Deep debugging. Chasing a subtle bug down through a chain of dependencies wants a stronger model.
• Close code review. Security holes. Slow code paths. Architectural anti-patterns. All of those want more horsepower.

Most of that routing is handled by smart switching on its own. If you pick a model yourself for something hard, Sonnet or Opus is the right call.

The jump up from older Haiku releases was real. Before 4.5, the Haiku line sat in its own lesser category, closer to a basic assistant than a smaller Sonnet. With 4.5, that gap closed far enough that, on simple work, most people couldn't spot a difference between a Haiku reply and a Sonnet one.

The quality floor also jumped. Older Haiku models would sometimes spit out code that was clearly wrong or half-done. Those failures got rare on the work Haiku 4.5 was built for.

Haiku 4.5 is still active. It is the current Haiku inside Claude Code. Point to it with the haiku alias. Or with the full model ID.

Related Pages

• All Claude Models for the full model timeline
• Sonnet 4.5 for the mid-tier everyday driver
• Opus 4.5 for top-tier reasoning
• Model selection guide for picking the right model per task
• Usage optimization for keeping Claude Code costs in check