Claude Haiku 5.5 targets the small, fast jobs that fill AI systems
token
A small piece of text counted by an AI.
benchmark
A test used to compare AI systems.
subagent
An AI given a smaller task by another model or application.
What happened
On October 7, 2026, Anthropic, the company that makes Claude, announced Claude Haiku 5.5. The official announcement presents it as a small model for high-volume, cost-sensitive work. Examples include summaries, organizing long conversations, database queries, classification, live customer support, browser use, and subagent work. Anthropic calls it the cheapest, fastest, and most capable small model it has released.
A token is a small piece of text counted by an AI. For requests up to 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. Above that size, the prices are $0.50 and $2.50. Anthropic says the model costs about 75% less to run on average than Haiku 4.5. Its footnote explains the calculation. Smaller requests cost 90% less, while larger requests cost 50% less.
The background
Anthropic positions Haiku 5.5 beside larger models, including Sonnet 5.5 and Opus 5.5. A larger model can handle the broad task. Haiku can handle a smaller task as a subagent. This arrangement may fit work that repeats many times. Haiku 5.5 is also the first Haiku model with an adjustable effort setting. Users can choose whether to favor cost or capability.
Why it matters
Small savings can matter when an AI service repeats a task thousands of times. Speed matters too. A faster model can reduce waiting in customer support or browser work.
Anthropic reports strong gains over Haiku 4.5 in several benchmarks. On the offline subset of OSWorld 2.1, Haiku 5.5 scored 72.4%, compared with 15.7% for Haiku 4.5. On Terminal-Bench 4.0, it scored 39.2%, compared with 0.0%. These tests cover computer use and command-line coding. Sonnet 5.5 still scored higher in the same table, at 83.9% and 70.6%. The announcement therefore does not present Haiku as a replacement for every larger model.
Anthropic also quotes early customer tests. Asana reported more than a 30% latency reduction and up to 2.5 times faster inference per agent turn. HubSpot reported a 92.8% average score on its CRM test. These are customer reports from early testing. They are not independent guarantees for every product.
What is confirmed
Haiku 5.5 is available through Claude and through Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic says its alignment results improved over Haiku 4.5. It also says its cybersecurity safeguards are stricter than Haiku 4.5. They still block penetration testing and other techniques more likely to help attackers.
What remains unknown
The announcement does not show how these results will translate to ordinary workplaces. We do not yet know the model’s long-term error rate, its real cost across mixed workloads, or how well teams will divide work between small and large models. The customer examples are early tests, not long-term studies.
What to watch next
The useful test will be real usage. Watch whether developers use Haiku for summaries, classification, and narrow tasks. Watch whether Sonnet or Opus remains necessary for complex agentic coding. The key question is not only whether Haiku is cheaper. It is whether the full system becomes cheaper without losing useful accuracy.
Attention is not proof
The story received 734 points and 373 comments on Hacker News. Those numbers measure community attention. They do not prove that Anthropic’s claims are correct.
Claude Haiku 5.5 handles many small jobs quickly
📰 Full story: Claude Haiku 5.5 targets the small, fast jobs that fill AI systems
Anthropic, the company that makes Claude, released a faster and cheaper model.
token
A small piece of text that an AI counts.
benchmark
A test for comparing AI systems.
subagent
An AI that handles a smaller task for another AI.
💡 The gist
- Claude Haiku 5.5 handles many small tasks.
- It aims to be fast and cheaper than Haiku 4.5.
- Hacker News showed interest, not proof of accuracy.
Anthropic announced Claude Haiku 5.5 on October 7, 2026. It is a smaller AI model. It can summarize text, organize long chats, search databases, and sort requests. It can also help larger Claude models.
These tasks may be small. They can happen many times. That makes speed and price important. Anthropic says Haiku 5.5 costs 90% less for requests up to 100,000 tokens. A token is a small piece of text. For larger requests, the price is 50% lower than Haiku 4.5.
Haiku 5.5 is not meant to replace every larger model. Sonnet 5.5 and Opus 5.5 can handle harder work. Haiku can take a smaller task from them. This role is called a subagent. It may help a system avoid using an expensive model for every step.
Anthropic published benchmark results. A benchmark is a test for comparing systems. In a computer-use test, Haiku 5.5 scored 72.4%. Haiku 4.5 scored 15.7%. In a coding test, the scores were 39.2% and 0.0%. Sonnet 5.5 still scored higher in those tests.
Anthropic also shared early customer results. Asana reported shorter waiting times. HubSpot reported a 92.8% average score on its own test. These results came from customer testing. They may not match every workplace.
The model has safety rules. Some harmful requests are blocked. It is available through Claude, Amazon Web Services, Google Cloud, and Microsoft Azure.
On Hacker News, the story received 734 points and 373 comments. That shows attention. It does not prove that every claim is correct.
💬 Haiku 5.5 is fast, but long jobs may cost more
The discussion split between people who see Haiku 5.5 as a fast helper for small tasks and people who think larger models or other pricing may win on complex work. The numbers below were reported by commenters or relayed from third-party tests.
- Some users reported that Haiku 5.5 was stronger than Luna and beat Sonnet 5 on parts of coding tests. These are user-reported results.
- In an image-to-HTML test, it struggled with a complex screen and passed the work to Opus 5.5, but it was reported to be fast and useful for tightly scoped tasks.
- The listed price was $0.10/MTok input and $0.50/MTok output up to 100,000 tokens, then $0.50 and $2.50. A commenter quoting the announcement said 90% of older Haiku 4.5 requests stayed within the lower range.
- Long agent jobs can exceed 100,000 tokens. Task-cost claims disagree: one discussion of the article said Haiku was cheaper than Luna, while a commenter citing Artificial Analysis said it was about three times more expensive.
- Reported speed was 93 tokens per second on average in one service and at least 137 in each Artificial Analysis test. That suggests a good fit for fast classification and summaries, but not an automatic winner for difficult jobs.
growing digest at 373 comments (revision 1). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
Claude Haiku 5.5 is an AI for small jobs
📰 Full story: Claude Haiku 5.5 targets the small, fast jobs that fill AI systems
Anthropic, the company that makes Claude, announced a new AI.
Anthropic
The company that makes Claude.
Claude Haiku 5.5
The name of Anthropic’s new AI.
Hacker News
A website where people discuss technology.
AI is a computer tool that works with words. Claude Haiku 5.5 can make short summaries. It can find words. It can sort messages.
Some small jobs happen again and again. Anthropic says Haiku 5.5 does them quickly. It also costs less than Haiku 4.5.
A bigger Claude model can handle harder jobs. Haiku 5.5 can handle smaller jobs beside it. This can save the bigger model for harder work.
Anthropic gave the AI safety rules. Some harmful requests are blocked.
Hacker News is a website about technology. The story received 734 points and 373 comments. This shows that many people noticed it. It does not show that the story is always correct.
💬 Haiku 5.5 is small and quick, but long requests can cost more
Haiku 5.5 got attention as a fast helper for short jobs. For harder or longer jobs, a bigger AI may still be better.
- Some people said it did better than Luna in certain tests. But when asked to build a difficult screen, it reportedly asked the bigger Opus for help.
- Short requests are cheaper: up to 100,000 tokens costs $0.10 for input and $0.50 for output. Longer requests cost $0.50 and $2.50. A commenter said the older Haiku's requests were short about nine times out of ten.
- The price of finishing one whole job was disputed. One result said Haiku was cheaper than Luna; another said it was about three times more expensive. So it may be good for quick sorting and summaries, but it is not always the best choice.
growing digest at 373 comments (revision 1). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.
💬 Haiku 5.5 looks fast and capable, but total cost depends on the job
Hacker News commenters saw Haiku 5.5 as a strong option for short, well-defined work and subagents, while others argued that long agentic tasks may make its pricing and cost per completed task less attractive. The performance and speed figures come from the article, commenter-reported tests, or third-party numbers relayed by commenters, not a settled overall verdict.
growing digest at 373 comments (revision 1). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.