Claude Haiku 5.5 explained: price, the 100K-token cliff, and when to use it

Updated November 2, 2026. Researched and drafted with AI help; every figure is checked against the linked source.

Anthropic released Claude Haiku 5.5 on October 7, 2026. It is the smallest Claude model, and Anthropic calls it the cheapest small model it has ever released, at $0.10 per million input tokens. It also has a pricing detail that is easy to miss: prompts over 100,000 tokens cost five times more. This guide covers what the model is, what it costs on real workloads, and where that 100K line bites.

The short version

Price: $0.10 input / $0.50 output per million tokens for prompts up to 100,000 tokens. Above that, $0.50 / $2.50.
Cache reads: $0.01 per million (or $0.05 above 100K). Batch API: 50% off.
Context: 1M tokens in, up to 128K tokens out (300K through the Batch API in beta).
API name: claude-haiku-5-5, on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS.
Previous Haiku: Claude Haiku 4.5 ($1 / $5), still listed on Anthropic’s pricing page. Anthropic says Haiku 5.5 costs about 75% less to run on average.
Matches: GPT-6 Luna exactly on list price for prompts up to 100K tokens.

What Anthropic says it is for

Anthropic calls Haiku 5.5 “the cheapest, fastest, and most capable small model we’ve ever released.” It points it at high-volume, cost-sensitive work: summarization, compaction, database queries and classification. It also suggests using it as a subagent under Claude Opus 5.5 or Claude Sonnet 5.5 in coding tools, and for speed-sensitive jobs such as live customer support and browser use.

Anthropic is also clear about the limits: it says Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding.

Two practical changes from Haiku 4.5:

It thinks. Haiku 5.5 is the first Haiku with an adjustable effort setting. Adaptive thinking is on by default and effort defaults to medium.
No sampling knobs. Anthropic’s docs say a non-default temperature, top_p or top_k returns a 400 error. If your code sets temperature today, remove it before switching.

The price, and the 100K-token cliff

All prices per million tokens, from Anthropic’s announcement and docs:

	Haiku 5.5, prompt up to 100K	Haiku 5.5, prompt 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 (5 min)	$0.125	$0.625	$1.25	$2.50

Anthropic says the price is 90% lower than Haiku 4.5 for prompts up to 100K tokens and 50% lower above that, and that about 90% of Haiku 4.5 requests fell under 100K.

The jump at 100K is steep. Here is what 1,000 requests with 1,000 output tokens cost on either side of the line:

Prompt size	Cost of 1,000 requests on Haiku 5.5
99,000 tokens	$10.40
101,000 tokens	$53.00

Two thousand extra input tokens make the bill about five times bigger. (This assumes the higher rate applies to the whole prompt, which is how Anthropic’s pricing table reads: it prices “prompts over 100,000 tokens”.) If your prompts hover near 100K (long documents, big retrieved contexts, agent histories), trimming or splitting them to stay under the line matters more than almost any other optimization. Other models have long-context tiers too, but at a higher threshold: Gemini 3.1 Pro charges more above 200K tokens and OpenAI’s GPT-6 models above 272K.

What 1,000 requests cost

We use the same three workloads as our cost per 1,000 requests guide, all well under 100K tokens:

Chat: 2,000 input and 500 output tokens.
Coding agent step: 2,000 fresh input tokens, 18,000 read from the prompt cache, 1,000 output tokens.
Summary: 10,000 input and 300 output tokens.

Model	Price in / out	1,000 chats	1,000 agent steps	1,000 summaries
Claude Haiku 5.5	$0.10 / $0.50	$0.45	$0.88	$1.15
GPT-6 Luna	$0.10 / $0.50	$0.45	$0.88	$1.15
GPT-5.6 Luna	$0.20 / $1.20	$1.00	$1.96	$2.36
DeepSeek V4.1 Flash*	$0.30 / $1.20	$1.20	$1.91	$3.36
Gemini 3.5 Flash-Lite	$0.30 / $2.50	$1.85	$3.64	$3.75
Gemini 3.8 Flash†	$0.75 / $3.75	$3.38	$6.60	$8.63
Claude Haiku 4.5	$1 / $5	$4.50	$8.80	$11.50
Claude Sonnet 5.5	$2 / $10	$9.00	$15.80	$23.00

* DeepSeek’s peak-hour price. Off-peak (all weekend, and weekdays outside 01:00–04:00 and 06:00–10:00 UTC) it is half.

† Promotional price through December 31, 2026. Google lists $1.50 / $7.50 from January 1, 2027.

On paper Haiku 5.5 and GPT-6 Luna tie for the cheapest model we track, at a tenth of what Haiku 4.5 cost for the same token counts.

The tokenizer caveat. Anthropic’s docs say Haiku 5.5 uses the newer Claude tokenizer, which counts the same text as roughly 30% more tokens than Haiku 4.5 did. Anthropic’s average figure, about 75% cheaper, is lower than the 90% list-price cut, and Anthropic says that figure accounts for the new tokenizer. The tokenizer also means a tie on list price is not necessarily a tie on your bill: if the same text came to 30% more tokens on Haiku 5.5 than on another model, its $0.45 per 1,000 chats would become about $0.59. Run a sample of your own prompts through both and compare the token counts in the API responses.

Long prompts: where GPT-6 Luna pulls ahead

Above 100K tokens, the tie breaks. Take 1,000 requests with a 150,000-token prompt and 1,000 output tokens, no caching:

Model	Cost of 1,000 requests
GPT-6 Luna	$15.50
Claude Haiku 5.5	$77.50
Claude Haiku 4.5	$155.00
Claude Sonnet 5.5	$310.00

GPT-6 Luna stays on its normal price up to 272K tokens, so for this job it costs a fifth of Haiku 5.5. Haiku 5.5 is still half the price of Haiku 4.5 here. Caching changes the picture again: a cached long document is read at $0.05 per million on Haiku 5.5 above 100K, so repeated questions over the same file stay cheap. Our prompt caching guide shows how much that saves.

Benchmarks, as reported by Anthropic

These are Anthropic’s own numbers from its announcement, not independent tests. GPT-6 Luna and Sonnet 5.5 figures are as Anthropic reported them.

Benchmark	Haiku 5.5	Haiku 4.5	GPT-6 Luna	Sonnet 5.5
OSWorld 2.1, offline subset (computer use)	72.4%	15.7%	48.9%	83.9%
Terminal-Bench 4.0 (agentic coding)	39.2%	0.0%	16.4%	70.6%
FrontierCode 1.1 Main (agentic coding)	46.4%	n/a	42.4%	52.1% (Xhigh effort)
Humanity’s Last Exam, no tools	45.9%	10.2%	n/a	56.9%
Chartography, no tools (visual reasoning)	46.4%	6.4%	29.1%	61.6%

Anthropic also published early customer reports. AlphaSense says Haiku 5.5 scored 0.84 against 0.76 for Haiku 4.5 on 400 “Ask in Document” queries, and Box says it scored 11 points higher than Haiku 4.5 at about half the latency.

The pattern in Anthropic’s own table is consistent: a large step up from Haiku 4.5, ahead of GPT-6 Luna on the tests where both are listed, and still clearly behind Sonnet 5.5, most of all on Terminal-Bench. Treat these as a starting point and check how the model does on your own tasks. Once people start using it, the Haiku 5.5 page shows live community votes on whether it feels better or worse on a given day, next to official incidents.

Also announced: a cheaper Sonnet 5.5 cache, and API credits for Max

Two other changes came with the launch:

Sonnet 5.5 cache reads were halved from $0.20 to $0.10 per million tokens. Anthropic says this cuts costs on most agentic tasks by about 20%. In our agent-step workload, 1,000 steps on Claude Sonnet 5.5 now cost $15.80 instead of $17.60 (about 10% less in this example), the same as GPT-6.1 Sol. More in our Sonnet 5.5 vs GPT-6.1 Sol guide.
Monthly API credits for Max, Ultra and Team plans. Anthropic’s help center says Max includes $100 and Ultra (formerly Max 20x) $200 of Claude API credits a month, and Team plans $20 per Standard seat and $100 per Premium seat, pooled and capped at $500. The credits work with any Claude model on the Claude Platform (not through Amazon Bedrock, Google Vertex AI or Microsoft Foundry), are claimed by linking a Console organization, don’t roll over and don’t raise your app usage limits. At Haiku 5.5’s prices, $100 covers about 222,000 of the chat requests above. See every plan on our plans page.

Should you switch?

On Haiku 4.5 today: yes for most workloads. The list price is 90% lower for prompts under 100K, and Anthropic’s own benchmarks show a large jump. Remove any temperature, top_p or top_k settings first, and recheck token counts because of the new tokenizer.
Choosing between Haiku 5.5 and GPT-6 Luna: under 100K tokens the list prices are identical, so pick on quality and real token counts for your own prompts. Above 100K tokens without caching, GPT-6 Luna is far cheaper.
Running a coding agent: keep Opus 5.5 or Sonnet 5.5 as the main model, as Anthropic recommends, and try Haiku 5.5 for subagent jobs like search, summaries and file reading.
Prompts near 100K tokens: trim or split them to stay below the line.

Put your own token counts into the cost calculator, compare it head to head with GPT-6 Luna, and if you run a site or README about it, a live rating badge is on the badges page.