Anthropic released Claude Haiku 5.5 on Wednesday at $0.10 per million input tokens for prompts up to 100,000 tokens, the same rate OpenAI charges for GPT-6 Luna. The rate is 90% below Haiku 4.5 on those prompts, but Anthropic’s own estimate of the average saving is about 75%, once a tokenizer that counts roughly 30% more tokens for the same text is included. Anthropic is aiming it at high-volume work, including live customer support and browser use, and at subagent tasks handed down by Opus 5.5 and Sonnet 5.5.

What Changed

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Pricing changes above 100,000 tokens

Output costs $0.50 per million tokens, compared with Haiku 4.5’s $5. Haiku 4.5’s input rate is $1 per million tokens.

Above 100,000 tokens, every rate multiplies by five. Input costs $0.50 and output $2.50 per million tokens. Cache reads rise from $0.01 to $0.05 per million tokens. Haiku 5.5 is the only current Claude model priced by prompt length.

Anthropic says about 90% of Haiku 4.5’s requests used prompts under 100,000 tokens. The new model’s context window grows from 200,000 to one million tokens, while its tokenizer counts approximately 30% more tokens for the same text. An existing prompt below the threshold may cross it after migration. The migration guide lists breaking changes, including errors for manual extended thinking and non-default sampling parameters.

Higher effort consumes more tokens

At maximum effort, Haiku 5.5 used roughly 162,000 output tokens per Intelligence Index task in Artificial Analysis’s launch testing, about three times GPT-6 Luna’s roughly 50,000 at the same output price, and more than Opus 5.5 at maximum effort.

At high effort, Haiku scored 38 using about 55,000 output tokens per task. Luna reached the same score at maximum effort with about 50,000. Output tokens are billed, so matching per-token prices does not mean matching cost per task.

Artificial Analysis’s cost figures do not yet reflect the tiered pricing. It plans cost-per-task coverage once its site supports tiered pricing.

Anthropic's numbers and an independent test

The launch benchmark table and customer testimonials are supplied by Anthropic, and Artificial Analysis is so far the only independent tester.

At medium, the default effort setting, Haiku 5.5 scored about 20% in Anthropic’s launch chart on Terminal-Bench 4.0, which measures coding and terminal tasks. At maximum effort, it scored 39.2%, against Luna’s 16.4% and Sonnet 5.5’s 70.6%.

Artificial Analysis’s independent testing, available at launch, put Haiku at 33% on the same benchmark, against 20% for Gemini 3.8 Flash and 13% for Luna. Anthropic says Sonnet 5.5 and Opus 5.5 “remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0.”

Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.

On the firm’s Intelligence Index, Haiku 5.5 reached 43 at maximum effort, up from Haiku 4.5’s 17 a year earlier. That placed it five points above Luna’s 38, one point behind Moonshot AI’s Kimi K3 and 13 behind Sonnet 5.5 at maximum effort.

Haiku 5.5 was the least accurate of the three on Artificial Analysis’s factual-knowledge test yet hallucinated least, because it was more willing to admit it did not know. On AA-Omniscience in the firm’s launch evaluation, its factual accuracy was 36%, below Gemini 3.8 Flash’s 55% and Luna’s 44%. Its hallucination rate was 40%, against 55% for Gemini and 77% for Luna.

Availability and developer changes

Haiku is available on the Claude Platform, AWS, Google Cloud, Microsoft Azure and GitHub Copilot. Anthropic added computer-use and browser-use support in beta to its Python and TypeScript SDKs.

Wednesday’s launch also halves Sonnet 5.5 cache-read prices from $0.20 to $0.10 per million tokens, which Anthropic estimates makes most agentic work on Sonnet 5.5 about 20% cheaper. Monthly API credits roll out this week: $100 for Max 5x, $200 for Max 20x and up to $500 pooled for Team subscribers.

One weak result

Haiku scored 35% on AutomationBench-AA, against 53% to 60% for GPT-6 Luna, Gemini 3.8 Flash and GLM-5.3 Flash. Artificial Analysis says a safety refusal issue during pre-release testing caused the model to over-refuse. Anthropic is working on a fix, and Artificial Analysis will re-run the test once it lands.

Frequently Asked Questions

How much does Claude Haiku 5.5 cost?

For prompts up to 100,000 tokens, $0.10 per million input tokens and $0.50 per million output tokens. Above 100,000 tokens, input costs $0.50 and output $2.50 per million tokens, and cache reads rise from $0.01 to $0.05.

Is Haiku 5.5 cheaper than GPT-6 Luna?

Not per token. Its short-prompt rates match GPT-6 Luna. Artificial Analysis found Haiku 5.5 used roughly 162,000 output tokens per Intelligence Index task at maximum effort, about three times Luna's roughly 50,000, so cost per task can differ.

Why does Anthropic estimate a 75% saving instead of 90%?

The 90% cut applies to prompts up to 100,000 tokens. Anthropic's 75% average also accounts for request sizes and a new tokenizer that counts approximately 30% more tokens for the same text than Haiku 4.5.

What is Haiku 5.5 good at?

Anthropic recommends it for summaries, compaction, classification, database queries and subagent tasks. In Anthropic's tests it scored 72.4% on the OSWorld 2.1 offline subset, against Luna's 48.9%. Anthropic recommends Sonnet 5.5 or Opus 5.5 for complex agentic coding.

Where is Claude Haiku 5.5 available?

On the Claude Platform, AWS, Google Cloud, Microsoft Azure and GitHub Copilot. Anthropic also added computer-use and browser-use support in beta to its Python and TypeScript SDKs.

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Claude Opus Plans Buy 5x the API Value of OpenAI's Sol, SemiAnalysis Tests Show
Andrew Megalaa’s subscription tests needed a book. For the SemiAnalysis study bearing his byline, analysts fed chunks of War and Peace into paid Claude and ChatGPT accounts, watching usage meters tick
The Implicator
Anthropic Investors Eye $2 Trillion IPO as Fable 5 Spending Lags GPT-5.6 Sol
Anthropic investors expect an October 2026 IPO at a valuation of at least $2 trillion, even as early spending data points to caution among companies buying its premium Fable 5 model. In Ramp’s July 20
The Implicator
Commerce Department Clears OpenAI's GPT-5.6 for Broad Public Release
The U.S. Department of Commerce has cleared OpenAI's GPT-5.6 model family for broad public release, a person familiar with the matter told Axios on Tuesday. OpenAI announced late Tuesday night that th
The Implicator
AI News

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai