A.I.
Rycode

Jeffrey 5 min read

Frontier models

Claude Haiku 5.5

Claude Haiku 5.5 is the third model in Anthropic’s Claude 5.5 family, released 7 October 2026. It is built for fast, high-volume work such as classification, routing, extraction, summaries, and subagent tasks. Anthropic says it is its fastest model to date and costs around 75% less to run than Claude Haiku 4.5 on average. The API id is claude-haiku-5-5. Token prices start at $0.10 input and $0.50 output per million, with cache reads at $0.01 per million.

Status
Current
AA score
43 (max effort)Artificial Analysis Intelligence Index, one independent score across many tests, fetched 8 Oct 2026
Price
$0.10 in, $0.50 outUS dollars per million tokens (word pieces) in and out, unless noted
Released
7 Oct 2026
Last checked
8 Oct 2026newest source date on this page

Training

The system card says training used a proprietary mix of public web data, public and private datasets, user data that users allowed for training, and synthetic data from other models, with cleaning steps that include deduplication and classification. The models overview lists a reliable knowledge cutoff and a training-data cutoff of June 2026.

Inference

Adaptive thinking is on by default. Haiku 5.5 is the first Haiku model with the effort setting, so users can trade cost for intelligence. The Claude API default is medium. The models overview lists it as the fastest model in the current lineup. Anthropic notes that its Opus models in fast mode still run faster. It uses the tokenizer Anthropic introduced with Claude Opus 4.7, so the same text counts as about 30% more tokens than on Haiku 4.5. Requests should leave temperature, top_p, and top_k at their defaults, because any other value returns an error.

Context

Context window is 1 million tokens. Max output is 128K tokens on the synchronous Messages API. On the Message Batches API, up to 300K output tokens is available with the output-300k-2026-03-24 beta header. Inputs are text and images. Output is text.

Methods

Anthropic aims Haiku 5.5 at quick, repeated work: summaries, context compaction, database queries, classification, live customer support, and browser use. It pairs with Opus 5.5 and Sonnet 5.5 as a subagent on coding work. Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding. Cyber safeguards are narrower than on Sonnet 5.5: they allow a wider range of defensive security work but still block penetration testing and other attacker techniques. Biology safeguards are the same as Sonnet 5, Sonnet 5.5, and Opus 5. Requests these safeguards block do not fall back to another model. Thinking blocks work only in the account that produced them, or in an account linked to it.

Benchmarking

Artificial Analysis gives Haiku 5.5 a score of 43 on its Intelligence Index at max effort (fetched 8 Oct 2026). The Intelligence Index is one number Artificial Analysis builds by running the same ten tests on every model. That ranks Haiku 5.5 second of 182 models in its price class. For scale, Claude Sonnet 5.5 scores 56 (Artificial Analysis, Claude Sonnet 5.5). GPT-6 Luna, which has the same $0.10 and $0.50 list price, scores 38 (Artificial Analysis, GPT-6 Luna).

Numbers below are Anthropic’s Haiku 5.5 announce table. Anthropic ran these tests itself, and no outside group has confirmed them yet.

Eval Haiku 5.5 Haiku 4.5 GPT-6 Luna Sonnet 5.5
Terminal-Bench 4.039.2%0.0%16.4%70.6%
FrontierCode 1.1 (Main)46.4%-42.4%52.1% at Xhigh
OSWorld 2.1 (offline subset)72.4%15.7%48.9%83.9%
Humanity's Last Exam (with tools)57.4%18.7%-64.5%
Humanity's Last Exam (no tools)45.9%10.2%-56.9%
GDPval-AA v2.1 (Elo)162073514371840
AA-Briefcase v1.1 (Elo)157861413361824
Chartography (no tools)46.4%6.4%29.1%61.6%

OSWorld here is version 2.1 on the offline subset, not the OSWorld 2.0 prints on the Opus 5.5 and Fable 5.1 pages. Anthropic’s announce charts also show Haiku 5.5 scores and cost at each effort level on OSWorld, GDPval-AA, and Humanity’s Last Exam.

Artificial Analysis measured about 243 output tokens per second at max effort (AA Haiku 5.5 model page, fetched 8 Oct 2026).

Hardware

Anthropic has not said which chips train or serve Haiku 5.5. Across the company, it trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs (Anthropic, Google and Broadcom compute deal). Its Google TPU deals are on the Ironwood page.

Access and pricing

Haiku 5.5, prompts up to 100K Haiku 5.5, prompts over 100K Haiku 4.5
Input / MTok$0.10$0.50$1
Output / MTok$0.50$2.50$5
Cache reads / MTok$0.01$0.05$0.10
Cache writes, 5 minutes / MTok$0.125$0.625$1.25
Cache writes, 1 hour / MTok$0.20$1-

The price depends on prompt length. Prompts up to 100,000 tokens get the low rate. Longer prompts cost five times as much per token. Cache reads are the cheap rate for reusing a prompt the model has already seen. The Batch API is 50% off these prices.

The list price is 90% below Haiku 4.5 on prompts up to 100,000 tokens. Anthropic says those short prompts made up about 90% of Haiku 4.5 requests. Above 100,000 tokens, the cut is only 50%. Add the new tokenizer, which counts the same text as about 30% more tokens, and Anthropic’s own average comes out at "around 75% less to run" than Haiku 4.5, not 90%.

Artificial Analysis shows the other half of the story, in figures fetched 8 Oct 2026. Running its full test set cost $0.21 per task on Haiku 5.5 at max effort, against $0.07 for GPT-6 Luna at the same list price. The reason is length. Haiku 5.5 wrote 440 million output tokens across the tests, against a median of 100 million for models in its class. The same sticker price does not mean the same bill. What you pay is price per token times the tokens a model actually writes.

Available on the Claude API, Amazon Bedrock (anthropic.claude-haiku-5-5), Google Cloud, Microsoft Foundry, and Claude Platform on AWS. In Claude Code, the haiku shortcut points to Haiku 5.5 only on the Anthropic API, and you need Claude Code v2.1.293 or later. On Claude Platform on AWS, Bedrock, Google Cloud, and Foundry, the shortcut still points to Haiku 4.5, so you pick claude-haiku-5-5 by name. Anthropic will not retire the model before 7 October 2027.

In the same launch Anthropic halved cache reads on Claude Sonnet 5.5, from $0.20 to $0.10 per million tokens. It says that makes Sonnet 5.5 about 20% cheaper on most agent work. Anthropic also added a monthly API credit: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team plans.

Sources