Claude Haiku 5.5: 90% Cheaper, One Catch

Claude Haiku 5.5 is Anthropic's lightweight model, released October 7, 2026. The official pricing landed at $0.10 input and $0.50 output per million tokens, exactly one-tenth of the previous Haiku 4.5 at $1/$5.

Most coverage stops at "90% cheaper." That price only holds when a single request stays under 100,000 tokens. Above that it jumps five times. Understand this boundary before you commit.

Pricing

Condition	Input	Output	Cache write	Cached read
Up to 100K tokens per request	$0.10	$0.50	$0.50	$0.05
Over 100K tokens per request	$0.50	$2.50	—	—

The previous Haiku 4.5 was $1/$5 with no short-versus-long split. Haiku 5.5 drops short requests to a tenth and long requests to half.

It directly matches GPT-6 Luna at $0.10/$0.50, but Luna has no 100K-token cliff.

How the 100K-token cliff catches you

A hundred thousand tokens sounds like a lot. Here is where it gets eaten:

One long document analysis with the document itself at 40K tokens plus a system prompt and conversation history crosses it easily
A twenty-turn conversation keeping all history
Batch processing several pages of contracts in one call

Crossing the line reprices the entire request, not just the portion above 100K. A 101K-token request costs five times what a 99K-token request costs. One line, all-or-nothing.

Short versus long request math

Short query (1K input + 0.5K output)

Under 100K: roughly $0.00035
1000 per day: $0.35
One month: roughly $10

Long document analysis (50K input + 10K output)

Over 100K: input at $0.50 x 0.05 = $0.025, output at $2.50 x 0.01 = $0.025, total $0.05
100 per day: $5
One month: roughly $150

Short queries are practically free. Long requests at the five-times rate cost about the same as Sonnet 5.5 at its flat $2/$10, but with less capable reasoning. That math does not hold up.

When to use it

Use it for:

Batch classification, tagging, sentiment (tens to hundreds of tokens each)
Simple FAQ auto-responses
Lightweight extraction (names, dates, amounts)
Short-turn chat assistant replies

Skip it for:

Full-length document reading in one pass
Multi-turn conversations keeping all history
Tasks requiring the model to read a lot of context before deciding

If you are not sure whether a request crosses 100K, run one pass and check the usage field.

Haiku 5.5 versus GPT-6 Luna

	Haiku 5.5	GPT-6 Luna
Official input	$0.10	$0.10
Official output	$0.50	$0.50
Long-request pricing	Over 100K -> 5x	Over 272K has a separate tier
Context	1M	1.05M
Protocol	Messages + Chat	Responses + Chat

Short requests are close. Long requests favor Luna for price stability. Run your own typical workload through both.

FAQ

Is it really a tenth of the old price?

For short requests, yes. Haiku 4.5 was $1/$5. Haiku 5.5 under 100K tokens is $0.10/$0.50. Long requests are $0.50/$2.50, roughly half.

How do I know if I will cross 100K?

Run one typical request and check the total tokens in the usage response. Add your expected context length.

Compared to Sonnet 5.5?

Sonnet 5.5 is $2/$10 and significantly more capable. If the task needs reasoning, the price gap is not worth the capability drop.

Available on HiAPI?

Not yet. It will appear in the model catalog when available.

How do I save money?

Cache system prompts and recurring context. Cached reads at $0.05 are practically free. Short requests plus caching is the cheapest combination.

Bottom line

The 90% price drop is real but only holds for short requests. Batch lightweight tasks on Haiku 5.5 at fractions of a cent each. Keep long documents away. GPT-6 Luna offers the same price without the cliff, so if your workload skews long, Luna is the safer choice.

Sources