Claude Haiku 5.5 cuts prices 90%. The real saving is 75%, and old code can break.

Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output tokens, a 90% cut from Haiku 4.5's $1 and $5. That puts Anthropic's smallest model at exactly the price of OpenAI's GPT-6 Luna, and it arrives with a 1 million token context window (up from 200K) and 128K max output. We have added it to our model database.

The headline number is real, but it is not the number that will show up on your invoice. Two details change the math, and a third can break your code.

The real saving is closer to 75%

Haiku 5.5 uses Anthropic's newer tokenizer, which Anthropic's docs say produces about 30% more tokens than Haiku 4.5 for the same text. You pay a tenth of the price per token but send more tokens, which is why some coverage quotes Anthropic as saying "around 75% less". That is still a large cut. Budget with 75%, not 90%.

The second detail is a price step. The $0.10/$0.50 rate applies to prompts up to 100,000 tokens. Above that, Haiku 5.5 costs $0.50/$2.50, so long-document workloads that use the new 1M context will pay the higher rate.

Per 1M tokens	Claude Haiku 5.5	Claude Haiku 4.5	GPT-6 Luna
Input (prompts up to 100K)	$0.10	$1.00	$0.10
Output	$0.50	$5.00	$0.50
Input above 100K tokens	$0.50	$1.00	Higher above 272K (reported)
Cache read	$0.01	$0.10	-
Context window	1M	200K	-

Batch processing halves the Haiku 5.5 rates again, to $0.05 and $0.25. GPT-6 Luna's pricing above comes from VentureBeat's reporting. Our AI Pricing Index tracks both.

Is it any good?

Anthropic's own benchmarks show a model that is far more capable than the one it replaces, particularly at agentic work. On OSWorld 2.1, which measures operating a computer, Haiku 5.5 scores 72.4% against Haiku 4.5's 15.7% and GPT-6 Luna's 48.9%. On Terminal-Bench 4.0 it scores 39.2% against Luna's 16.4%.

Independent testing adds a useful caveat. Artificial Analysis scores Haiku 5.5 between 29 and 43 on its Intelligence Index depending on the effort setting, against 38 for GPT-6 Luna. At its maximum effort it beats Luna, but it also uses about three times as many output tokens per task to get there. Cheap per token is not the same as cheap per job: turn effort down for simple tasks.

"We saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn." - Aaron Vinh, Staff Software Engineer, Asana

Check your code before you switch

This is not a drop-in replacement for Haiku 4.5. According to Anthropic's migration guide, these changes return errors or quietly break older code:

1. Thinking is on by default. Responses can start with a thinking block, so code that reads the first content block as the answer gets nothing. Read blocks by type.
2. Thinking tokens count toward max_tokens. A tight limit can end the response before any text appears. Raise small limits or switch thinking off for simple tasks, which is allowed at the default effort level.
3. Sampling parameters are gone. Custom temperature, top_p or top_k values return a 400 error.
4. Assistant prefill returns a 400 error. Use structured outputs or tools to control format instead.

We ran into the first two ourselves. When we moved our own content pipeline to a newer Claude model in September, the first-block assumption silently broke article generation for a day. Before switching this site's automation to Haiku 5.5 today, we audited every call for all four issues. If you are migrating, do the same check first, then measure token counts on a sample of real requests before projecting savings.

Who should switch now

If you run high-volume, simple calls (classification, extraction, routing, short summaries) on Haiku 4.5, switching is close to free money once your code passes the checklist above. If you use a mid-tier model for agent sub-tasks, Haiku 5.5's agentic scores make it worth testing as a much cheaper replacement. If you are on GPT-6 Luna, the two now cost the same, so the decision comes down to which one performs better on your workload. For apps rather than API, our ChatGPT vs Claude comparison covers the products you subscribe to.