Thursday, October 8, 7:30 AM
Every night the business collects emails, website inquiries and Google reviews. Until now the agent only read what was flagged as urgent, because reading everything with a strong model cost more than it was worth. Last night Anthropic released Claude Haiku 5.5, and the per-token price of its small model dropped by 90%.
The question is no longer "what is worth reading?" It is "why not read everything?"
Claude Haiku 5.5 is Anthropic's small, fast, cheap model, released on October 7, 2026. In practice this changes the economics of AI agents: when a model that can classify, summarize and check costs a tenth of what it cost a week ago, the decision about which tasks are worth handing to an agent changes. We checked every figure against the official pages and marked which numbers come from Anthropic itself.
In short
- Claude Haiku 5.5 is Anthropic's small model for high-volume work: classification, summarization, data extraction and subagents. Released October 7, 2026.
- The price: $0.10 per million input tokens and $0.50 for output, for requests up to 100K tokens. Above that threshold, 5 times more. That is 90% less per token than Haiku 4.5.
- For the first time in the Haiku line there is an effort level: five settings that trade savings against quality. The default is medium.
- What it means for a business: small, repeating tasks that did not pay off for an agent, like reading every email and every review, become very cheap. Long, complex tasks still belong to larger models.
What Claude Haiku 5.5 is, and what changed since Haiku 4.5
Haiku 5.5 is the smallest of Anthropic's four current models, alongside Sonnet 5.5, Opus 5.5 and Fable 5.1. In its comparison table it is marked as the fastest, built for high-volume tasks that need a quick answer. If the term "model" is still fuzzy, we wrote about what an LLM is.
Sources: Anthropic's announcement page, the model page and the "What's new" page in the documentation, October 8, 2026. The 75% figure is Anthropic's, not our own measurement.
What else is new, according to the documentation:
- Adaptive thinking. The model decides for itself when and how much to think, and the effort level steers it.
- A new tokenizer. The same text is counted as about 30% more tokens than in Haiku 4.5. We come back to this in the calculator.
- Availability. Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. According to Anthropic, also in Claude.ai on every plan, including the free one. The model name in the API:
claude-haiku-5-5.
The chef doesn't peel onions
The simplest way to see why this announcement matters is a restaurant kitchen.
The chef is the big model: Opus or Sonnet. The chef plans the menu, tastes and decides. The chef's hour is expensive, and nobody wants it spent peeling onions.
The prep crew is the small model. It chops, sorts and preps, a lot and fast. Claude Haiku 5.5 is a more skilled prep crew than the last one, and it costs a tenth as much per token.
When that happens, the way you work changes: it makes sense to prep everything ahead, check every dish, and leave only the decisions to the chef.
Why a cheap model changes the economics of agents
An AI agent doesn't send the model one request but dozens: it reads, classifies, checks, summarizes and fixes. Most steps are simple. When every step runs on a large model the cost climbs, so steps get skipped: only some of the emails are read, the website is checked once a week instead of every day.
Anthropic itself positions Haiku 5.5 as a subagent alongside Opus 5.5 and Sonnet 5.5. In practice it looks like this:
An agent with a crew: the big model plans, Haiku executes, you approve
- You give a taskFor example: "go through all of this week's inquiries"
- The big model plansBreaks it into steps and decides what needs judgment
- The Haiku crew executesClassifies, summarizes and checks every item, many in parallel
- You approveYou get a summary and suggestions. Nothing sensitive happens without your approval
In gold: the steps where you are involved. An illustration of a common work structure, not a specific product.
So what suddenly becomes worth doing? Four kinds of work where we see the biggest change:
Read everything
- A summary of every incoming email, not just the urgent ones
- A digest of every review and every lead
Background checks
- A daily pass over the website: broken links, mismatched prices
- Comparing live ads against the website
Routing and sorting
- Where each inquiry goes: sales, service or you
- Extracting details from invoices and forms
Subagents
- Many small steps in parallel
- A second check of an output before it reaches you
If you are looking for ideas for such tasks, we have a list of 10 tasks worth handing to an agent.
How much Claude Haiku 5.5 costs, and what the 100K-token threshold is
What is new here: the price depends on the length of the request. According to the pricing page, a request whose input (prompt) is shorter than 100K tokens pays the low price, on both input and output. A longer request pays 5 times as much on everything. On Anthropic's other current models the price is flat across the entire context window.
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Claude Haiku 5.5, request up to 100K tokens | $0.10 | $0.50 |
| Claude Haiku 5.5, request over 100K tokens | $0.50 | $2.50 |
| Claude Haiku 4.5 | $1 | $5 |
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Luna (OpenAI) | $0.10 | $0.50 |
| Gemini 3.1 Flash-Lite (Google) | $0.25 | $1.50 |
| Gemini 3.8 Flash (Google) | $0.75 | $3.75 |
Sources: the official pricing pages of Anthropic, OpenAI and Google, as of October 8, 2026, standard price with no discounts. On Haiku 5.5, reading text that is already cached costs $0.01 per million tokens, and Anthropic's Batch API gives a 50% discount. The Gemini 3.8 Flash price is valid until December 31, 2026, and according to Google it doubles from January 1, 2027. On GPT-6 Luna, a request over 272K tokens is priced higher.
Output price per million tokens, in dollars
Source: the official pricing pages, October 8, 2026. A price comparison only, not a quality comparison. Each company counts tokens differently, so a price per token is not a price per task.
How much is 100K tokens? A token is a small piece of text in which the model counts its work (the full explanation is in what a token is). According to Anthropic, a million tokens in the new tokenizer is roughly 555K English words, which means 100K tokens is about 55K words. There is no official figure for Hebrew. Here are the examples Anthropic gives on the pricing page:
The key point for agents: an agent's request grows over time. At every step the instructions, the history and the tool results are sent again, and an agent that works for an hour on one task can cross the threshold without anyone noticing. So the right pattern with Haiku 5.5 is short tasks, each starting clean. That is exactly what subagents do.
Calculator: what a repeating task will cost you
Choose how many times a month the task runs, and how much text goes in and out each time. The calculator compares Haiku 4.5 with Haiku 5.5 using the official prices.
Monthly cost calculator: Haiku 4.5 vs. Haiku 5.5
13.50 dollars a month on Haiku 4.5
1.76 dollars a month on Haiku 5.5
In this scenario Haiku 5.5 is 87% cheaper than Haiku 4.5. Every request is under the 100K-token threshold. For comparison, Sonnet 5.5 would cost about $35.10.
An estimate only. The tokens on the sliders are counted as in Haiku 4.5. Excludes caching, VAT and tool fees. On Haiku 5.5 thinking is counted as output, so the actual output depends on the effort level. The real number comes from measuring your own tasks.
Try dragging the input above 77K tokens, which is about 100K in the new tokenizer, and watch the savings shrink. Anthropic itself cites savings of 90% on short requests, 50% on long ones, and about 75% on average.
What effort level means, in plain English
Haiku 5.5 is the first Haiku with an adjustable effort level. Think of the difference between "take a quick look" and "check it thoroughly." According to Anthropic's documentation, effort affects everything the model produces: how much it thinks, how many actions it takes with tools and how much detail it gives. Lower effort is faster and cheaper, with some drop in capability.
| Level | What Anthropic recommends for Haiku 5.5 |
|---|---|
| low | The cheapest and fastest: chat, short tasks with tools, simple high-volume requests |
| medium | The default. The starting point for most work, including agent code |
| high | Professional work, longer agent tasks, closer adherence to instructions |
| xhigh / max | Only if a test on your own tasks shows an improvement. And then it is worth comparing against Sonnet 5.5 too |
And a warning from the documentation: with long agent instructions at low, the model may skip searching, stop early or skip a check. How much does it matter? The system card has a per-level measurement on a professional medical benchmark:
Haiku 5.5 on PhysicianBench: success rate by effort level
Source: the Claude Haiku 5.5 system card, Anthropic, October 7, 2026. Time per task: about 48 seconds at low, about 133 seconds at high and about 11 minutes at max. A vendor test, in one domain.
The same picture repeats in a professional-work benchmark called GDPval-AA. According to the system card, Haiku 5.5 scored 1620 at max, but at the default level (medium) it scored 1277, using about a tenth of the output tokens. The principle: effort is a dial, not a constant. For sorting inquiries, low or medium is enough. When a mistake is expensive, a larger model is sometimes better than max on a small one.
How good is it really? Anthropic's numbers
All the numbers in the table were published by Anthropic, in the system card and on the announcement page, and Anthropic chose and ran some of the tests itself. In most tests Haiku 5.5 was evaluated at max effort, not the default. The GPT-6 Luna results are as reported by Anthropic: on OSWorld it ran the model itself through OpenAI's API, and on Terminal-Bench the figure is taken from the test's public leaderboard.
| Benchmark (what it tests) | Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 | GPT-6 Luna |
|---|---|---|---|---|
| GDPval-AA v2.1 (professional work, Elo score) | 1620 | 735 | 1840 | 1437 |
| OSWorld 2.1, offline subset (computer use, partial score) | 72.4% | 15.7% | 83.9% | 48.9% |
| Terminal-Bench 4.0 (terminal work) | 39.2% | 0.0% | 70.6% | 16.4% |
| Humanity's Last Exam, no tools (hard questions) | 45.9% | 10.2% | 56.9% | Not published |
| SWE-bench Multilingual (code fixing) | 83.7% | 67.4% | 90.3% | Not published |
| Global MMLU (knowledge in 42 languages) | 87.8% | Not published | 92.1% | Not published |
The jump from Haiku 4.5 is very large, especially in computer use and professional work. But Sonnet 5.5 still leads in every row, and in terminal work the gap is almost double. For a wider picture across all vendors, see our ranking based on independent benchmarks.
Which model for which task
Dividing the work between Claude models, based on Anthropic's definitions and our experience
Haiku 5.5
- Classifying and routing inquiries and leads
- Summarizing emails, reviews and short documents
- Extracting data from forms and invoices
- Subagents and background checks
- First-line chat that must answer fast
Sonnet 5.5
- Long agent tasks with many tools
- Content published in the business's name
- Campaign analysis and reports
- Mid-level code
Opus 5.5 or Fable 5.1
- Planning and coordinating a team of agents
- Complex documents and hard problems
- Complex code and tasks that run for hours
- Decisions where a mistake is expensive
The definitions for each model follow Anthropic's model table. Assigning business tasks to models is our opinion, not an Anthropic recommendation.
The Haiku line timeline
Haiku 4.5 is currently marked as Legacy. According to Anthropic, it will not be retired before October 15, 2026, and at least 60 days' notice will be given before retirement. No need to panic, but it is worth planning a move.
Five myths about Claude Haiku 5.5
Myth"90% cheaper per token, so the bill drops by 90%"
In practiceThe same text is counted as about 30% more tokens, thinking is counted as output, and requests over 100K tokens cost 5 times as much. Anthropic itself talks about roughly 75% on average.
Myth"A small model is a weak model"
In practiceIn Anthropic's tests, Haiku 5.5 beats GPT-6 Luna in several of them and comes close to Sonnet 5.5. But in terminal work and long tasks the gap to Sonnet is still large.
Myth"There's a million-token window, so I'll pour everything in"
In practiceAbove 100K tokens the price jumps 5 times. It is usually better to split the work into small pieces.
Myth"The max level is always better"
In practiceAnthropic recommends starting at medium and going up only if a test shows an improvement. At max a task can take minutes instead of seconds, and use several times as many tokens.
Myth"The benchmarks prove it fits my business"
In practiceThese are tests published by the vendor, most of them at max. They show a direction. The answer comes from ten or twenty of your own real tasks.
Where Claude Haiku 5.5 is not the right choice
Here is our opinion, based on what Anthropic itself publishes:
- Long, complex agent tasks. On Terminal-Bench the gap is 39.2% against 70.6% for Sonnet 5.5. When the agent has to plan dozens of steps and recover from mistakes, a large model saves more than it costs.
- Working on screen with untrusted content. Haiku 5.5 is far more resistant than its predecessor to malicious instructions hidden in content, known as prompt injection. In a Gray Swan test, the chance of an attacker succeeding within 15 attempts fell from 83.2% to 7.1%. That is still higher than Sonnet 5.5 (3.4%) and Opus 5.5 (1.0%), and most of the remaining vulnerability is in screen control: 24.4%. An agent that browses unfamiliar sites and can act needs safeguards and approvals, on any model. We went deeper in what Computer Use is.
- Very long documents. Above 100K tokens Haiku 5.5 is 4 times cheaper than Sonnet 5.5, not 20 times. When you need a deep understanding of a long document, the price difference doesn't always justify the quality difference.
- Text published in Hebrew in the business's name. Anthropic doesn't publish a score for Hebrew. On the 42-language knowledge test Haiku 5.5 scored 87.8%, against 92.1% for Sonnet 5.5. For an internal summary that is enough. For a post going out to an audience, test it first.
- Decisions where a mistake is expensive. Budget, pricing, an official reply to a customer. According to the system card, in the automated behavior audit it refused unnecessarily more than any other model tested, although in a different test it refused less than Haiku 4.5.
Our bottom line: Haiku 5.5 is an excellent team member, not the team's manager. It pays off when a large model plans, it executes many small steps, and you approve anything sensitive.
Is Haiku 5.5 right for your task?
Five questions about one task you are considering handing to an agent.
What this means for Platin Digital clients
Platin Digital is an OpenAI Select Partner, and the agents we build use whichever models and tools fit each business best, from OpenAI, Anthropic, Google and others. Still, this announcement matters to you too: Haiku 5.5 joins OpenAI's GPT-6 Luna, which has the same per-token price, and both show the same trend. Small, capable models turn tasks that did not pay off for an agent into worthwhile ones. If you are comparing platforms, we have a full comparison of ChatGPT, Gemini and Claude.
- The model's price is just one line on the bill. Setup, integrations and support are a large part of the cost. We broke it down in how much an AI agent costs a business.
- Choose the model by task. Small, repeating steps on a small model, decisions on a large model, with any vendor.
- The approval rules don't change. Publishing, budget, deletion and messages in the business's name keep waiting for your approval, on every model and at every price.
If you want to understand what such an agent does on an ordinary day in a business, start with the guide AI agent for business.
Frequently asked questions
What is Claude Haiku 5.5?
Claude Haiku 5.5 is Anthropic's small, fast language model, released on October 7, 2026. It is built for high-volume work such as classification, summarization, data extraction and subagents, costs from $0.10 per million input tokens and $0.50 for output, supports a context window of one million tokens, and is the first Haiku with an adjustable effort level.
How much does Claude Haiku 5.5 cost on the API?
Up to 100K tokens per request: $0.10 per million input tokens and $0.50 for output. Above that: $0.50 and $2.50. The Batch API gives a 50% discount. According to Anthropic, on average it is about 75% cheaper than Haiku 4.5.
What is the effort level, and which one should I choose?
The effort level sets how much the model thinks and works before it answers. Haiku 5.5 has five levels: low, medium, high, xhigh and max. The default is medium, and Anthropic recommends starting there, dropping to low for simple, fast tasks, and going up only if a test on your own tasks shows an improvement.
Can I use Haiku 5.5 in Claude.ai?
Yes. According to Anthropic, Free, Pro, Max, Team and Enterprise users can choose Haiku 5.5 in Claude.ai, in the browser, on iOS and on Android. It is also available in Claude Code, the API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Should I move from Haiku 4.5 to Haiku 5.5?
In most cases yes, but with testing. Haiku 4.5 is marked as Legacy, and Anthropic recommends considering a move. According to the "What's new" page, some changes will break existing code: temperature, top_p and top_k return an error, manual thinking with budget_tokens was replaced by the effort level, and the tokenizer counts about 30% more tokens. Run your tasks on both models before you switch.
Is Claude Haiku 5.5 good in Hebrew?
Anthropic doesn't publish a score for Hebrew. It says the model is multilingual and that quality varies between languages. On the 42-language knowledge test it scored 87.8%, less than Sonnet 5.5. For text you will publish, test it on real examples of your own.
Sources
- https://www.anthropic.com/claude-haiku-5-5
- https://www.anthropic.com/news
- https://www.anthropic.com/claude/haiku
- https://www.anthropic.com/document/claude-haiku-5-5-system-card
- https://platform.claude.com/docs/en/models/haiku-5-5/overview
- https://platform.claude.com/docs/en/models/haiku-5-5/whats-new-haiku-5-5
- https://platform.claude.com/docs/en/about-claude/models/overview
- https://platform.claude.com/docs/en/about-claude/pricing
- https://platform.claude.com/docs/en/build-with-claude/effort
- https://platform.claude.com/docs/en/about-claude/model-deprecations
- https://platform.claude.com/docs/en/models/haiku-4-5/overview
- https://www.anthropic.com/news/claude-haiku-4-5
- https://www.anthropic.com/news/claude-3-haiku
- https://www.anthropic.com/news/3-5-models-and-computer-use
- https://support.claude.com/en/articles/12138966-release-notes
- https://aws.amazon.com/blogs/machine-learning/introducing-claude-haiku-5-5-on-aws/
- https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/haiku-5-5
- https://developers.openai.com/api/docs/models/gpt-6-luna
- https://developers.openai.com/api/docs/pricing
- https://ai.google.dev/gemini-api/docs/pricing
