Claude Haiku 5.5 Is 20x Cheaper Than Sonnet. Here's Where I'd Actually Use It

Most model launches are about the top of the stack: the smartest model, the highest benchmark. This week's is about the bottom, and for anyone running agents it might matter more to your bill.

Claude Haiku 5.5 shipped on October 7, 2026 (claude-haiku-5-5). It's the first Haiku in a while, and the first Haiku with an adjustable effort setting. Anthropic's own pitch is high-volume, cost-sensitive work: summaries, compaction, database queries, classification, live support, browser use, and subagents that work alongside the bigger models.

That last one is the reason I'm writing this. If you build with Claude Code or the API, the question isn't "should I switch to Haiku." It's "which parts of my workflow don't need a genius."

The numbers

Prices per million tokens, from Anthropic's announcement. Haiku 5.5 is priced by prompt length. These are the rates for prompts up to 100K tokens (above that, input is $0.50 and output is $2.50).

Input is $0.10 and output is $0.50, against $2 and $10 for Sonnet 5.5. That's 20x cheaper per token. Cache reads drop to $0.01. Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average, which is smaller than the sticker gap because of a tokenizer change I'll get to below.

It isn't the model for your hard problems

Read the benchmark table before you get excited. On Terminal-Bench 4.0, the eval closest to what Claude Code does all day, Haiku 5.5 scores 39.2% against Sonnet 5.5's 70.6%. Anthropic says plainly that Sonnet 5.5 and Opus 5.5 stay better for complex agentic coding.

So I'm not moving my main Claude Code session to it. What I'm doing is splitting the work.

The pattern: a big model orchestrates, and Haiku does the high-volume reading. Exploring a codebase, skimming 40 files to find the three that matter, summarizing a long thread, tagging a pile of inbound messages. Those jobs eat enormous numbers of input tokens and need competence, not brilliance. At $0.10 per million input, you stop rationing them.

In Claude Code you can point a subagent at a specific model in its definition. If you haven't built one, my guide to the Claude Agent SDK covers the structure, and the routines post is the place to look for scheduled work that doesn't need a top-tier model. One caution: Anthropic's launch page doesn't mention Claude Code availability, so check your model picker before assuming Haiku 5.5 is wired in.

The effort dial

Haiku 5.5 is the first Haiku with an effort setting, and Anthropic's charts show five levels: low, medium, high, xhigh, max. Adaptive thinking is on by default, so the model decides when to think, and the docs say effort is the better lever for trading quality against speed and cost than switching thinking off.

My starting rule, which is a plan to test rather than a result: low for classification and routing, medium (the default) for summaries and extraction, and anything that needs high or above probably belongs on Sonnet anyway. I'll tell you what holds up once I've run real workloads through it.

The migration traps

If you have code pointed at Haiku 4.5, don't just swap the model string. From the what's new page, these will error or surprise you:

• Manual extended thinking errors. Replace budget_tokens with adaptive thinking.
• Sampling parameters can error. Drop temperature, top_p and top_k.
• Assistant prefill errors. End your messages with a user turn.
• Responses can start with a thinking block. Select content by type, not by position, or you'll read the wrong block as the answer.
• The same text counts as about 30% more tokens than on Haiku 4.5, because of the newer tokenizer. Recount your prompts and revisit max_tokens and cost estimates. This is why the real saving is nearer 75% than the 90% the price table implies.
• Earlier thinking blocks stay in context, so multi-turn conversations carry more input tokens. Clear old blocks if that hurts.

On the plus side, the context window is 1M tokens with up to 128K output, and the minimum cacheable prompt drops to 512 tokens from 4,096, so short, repeated prompts now cache.

What I'd do this week

1. List every place your agent or automation calls a model. Mark each one "needs judgment" or "needs reading."
2. Pick the single highest-volume "needs reading" job and run it on Haiku 5.5 at low or medium effort next to your current model.
3. Compare outputs on twenty real examples, not a benchmark. If they're equivalent, keep the savings.
4. Leave your main coding session alone.

One more note from the announcement: Max 5x gets $100 in monthly API credits, Max 20x gets $200, and Team gets up to $500 pooled, which is a free way to run that test.

This is the same lesson as every model release this fall: the interesting move isn't picking the best model. It's matching the model to the job. If you want help designing that split for your own workflow, that's most of what I do live in 1:1 coaching and private team training.