Claude Haiku 5.5: $0.10 per million tokens, but heavy token use

On 7 October 2026 Anthropic introduced Claude Haiku 5.5, which it calls the cheapest, fastest and most capable small model it has released. It is aimed at summaries, database queries, classification, live customer support and use as a subagent next to larger models. On average it should cost around 75 percent less than Haiku 4.5. The benchmarks in the announcement are vendor claims. An independent score arrived a day later.

Prices and vendor benchmarks come from Anthropic's announcement of 7 October 2026. Independent measurements come from The Decoder's update of 8 October (data from Artificial Analysis). (Quelle: Simple Icons, file anthropic.svg (CC0))

The prices

Per million tokens, according to Anthropic:

	Haiku 5.5 (up to 100k / over 100k)	Haiku 4.5	Sonnet 5.5
Input	$0.10 / $0.50	$1.00	$2.00
Output	$0.50 / $2.50	$5.00	$10.00
Cache read	$0.01 / $0.05	$0.10	$0.10
Cache write	$0.125 / $0.625	$1.25	$2.50

The cheap tier applies to prompts up to 100,000 tokens. Anthropic says those make up around 90 percent of requests to the previous Haiku. Above that limit prices go up fivefold, as The Decoder notes. Anthropic also halves the cache-read price for Sonnet 5.5. According to The Decoder, the context window grows from 200,000 to one million tokens.

Input tokens follow the same pattern: $0.10, $1.00 and $2.00. (Quelle: Anthropic, Claude Haiku 5.5, 7 October 2026)

A worked example

A classification job with 10 million requests, 2,000 input tokens and 100 output tokens each. That is 20 billion input and 1 billion output tokens. At list prices:

Model	Input	Output	Total
Haiku 5.5	$2,000	$500	$2,500
Haiku 4.5	$20,000	$5,000	$25,000
Sonnet 5.5	$40,000	$10,000	$50,000

This is our own arithmetic with assumed volumes and no caching. The tokenizer overhead is not included; neither Anthropic nor The Decoder gives a fixed figure for Haiku. For the Opus models, according to The Decoder, the new tokenizer alone raised token use by about 30 percent.

What Artificial Analysis measures

On 8 October Artificial Analysis rated the model. At the highest effort setting Haiku 5.5 scores 43 on the Intelligence Index, narrowly leading the small models. Sonnet 5.5 sits 13 points above, says The Decoder.

Sonnet 5.5 is 13 points above Haiku 5.5. All values are Artificial Analysis data as reported by The Decoder. (Quelle: The Decoder, update of 8 October 2026)

The price per token says little without the consumption. At the highest setting Haiku 5.5 needs about 162,000 output tokens per index task, GPT-6 Luna about 50,000. On "high", Haiku 5.5 scores 38 with around 55,000 tokens, and Luna gets the same score with 50,000. Going from "xhigh" to "max" buys two points for 1.8 times the consumption. Cost per task at index level is $0.21 for Haiku 5.5 (max); Luna reaches a similar score for about a third of that.

On the plus side is the hallucination rate: 40 percent for Haiku 5.5 against 77 percent for Luna. Factual knowledge is lower (36 against 44 percent on the AA-Omniscience test).

Who it pays off for

The low list price works best on short bulk tasks such as classification, extraction and summaries, where the gap to Haiku 4.5 is real. On tasks that need the highest effort the advantage shrinks, because consumption rises. That is our own judgment. We have no measurements for bulk tasks of this kind, and the index tests harder work.

For us this changes little about the split. Anything touching WinLine, banking or Logis data runs on local models and goes to no hosted API, even after this price cut. Our local production lane runs on four RTX 5090s. For uncritical bulk work the Haiku 5.5 API is a new point of comparison, but a reliable calculation needs the utilization and power costs of your own cards, which we have not worked out here. Anyone who wants to test should measure token consumption on their own tasks and calculate with those values instead of the list price.