Claude Haiku 5.5 — Anthropic’s cheapest and fastest model, released on October 7, 2026: $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens, a context window of up to 1 million tokens, and adjustable reasoning depth. It’s designed for high-volume routine work: classifying inquiries, extracting fields from documents, brief summaries, and serving as the execution layer under a stronger model. According to Anthropic, it costs about 75% less on average than Haiku 4.5. The key things businesses need to know before deploying it: pricing jumps sharply after the 100,000-token threshold, it’s better not to trust it with complex agentic code, and Anthropic’s service is not officially available from Russia.
This article was prepared by AIrassvet based on publicly available sources, verified on October 11, 2026. It is intended for business owners and executives, CIOs, and product leaders choosing a language model for high-volume tasks and budgeting for AI. We did not test the model ourselves: all benchmark results come from Anthropic and summaries in industry publications, while the cost calculations are our own examples based on published pricing. AIrassvet is not affiliated with Anthropic, OpenAI, or any other companies mentioned, and the section at the end describes the publisher’s services.
Contents
- What was released on October 7, 2026
- In Brief: Key Takeaways on Haiku 5.5
- Specs and Pricing
- The 100,000-Token Threshold: When a Cheap Model Gets Expensive
- How Much a Task Costs: Calculation Examples
- Benchmark Results and What They Mean
- Reasoning Depth Modes
- What Business Tasks Haiku 5.5 Is Good For
- The “planner and executor” setup
- When Haiku 5.5 Is Not a Good Fit
- Switching from Haiku 4.5: What to Check
- Availability in Russia and Data Requirements
- How to Test the Model on Your Own Tasks
- Frequently Asked Questions
- Primary Sources and Review Method
What was released on October 7, 2026
Anthropic introduced Claude Haiku 5.5 on the model page and called it the company’s cheapest, fastest, and most capable small model. The API identifier is claude-haiku-5-5. The model is available through Anthropic’s own platform as well as in Amazon Web Services, Google Cloud, and Microsoft Azure.
What is known as of October 11, 2026:
- Price. $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens. For longer requests: $0.50 and $2.50.
- Savings. According to Anthropic, typical tasks cost about 75% less on average than on Haiku 4.5. For short requests, the rate is 90% lower; for long ones, 50% lower.
- Context and output. Up to 1 million input tokens and up to 128,000 output tokens, according to DataNorth. It accepts text and images as input.
- Reasoning depth. For the first time in a Haiku-class model, there is an effort setting: from Low to Max. The default is medium.
- Side benefit: lower Sonnet 5.5 pricing. Cache reads for Claude Sonnet 5.5 are now half as expensive, dropping from $0.20 to $0.10 per million tokens; according to Anthropic, this reduces the cost of most agentic tasks by about 20%.
Positioning. Anthropic explicitly names the use cases: short summaries, context compression, database queries, classification, subagent work, real-time customer support, and browser actions. For complex agentic programming, the company still recommends Sonnet 5.5 and Opus 5.5.
In Brief: Key Takeaways on Haiku 5.5
- It’s a model for scale, not complexity. Its strength is thousands of repetitive short tasks with a clearly defined response format.
- Pricing depends on prompt length. Up to 100,000 tokens, the model is among the cheapest on the market; after the threshold, the entire request is billed at a rate five times higher.
- The improvement over Haiku 4.5 is substantial. According to Anthropic, in the GDPval-AA office-work benchmark, the score rose from 735 to 1,620, and in OSWorld 2.1 computer control, from 15.7% to 72.4%.
- Sonnet 5.5 is still clearly stronger. In agentic programming, Terminal-Bench 4.0 shows an almost twofold gap: 39.2% versus 70.6%.
- Token counts will be slightly higher. The new tokenizer counts the same text slightly differently, so you’ll need to recalculate the budget using your own data.
- For Russia, this is not a ready-made option. Direct access to Anthropic from Russia is not supported; companies need a lawful setup or an alternative.
Specs and Pricing
| Parameter | Claude Haiku 5.5 | Claude Haiku 4.5 | Claude Sonnet 5.5 |
|---|---|---|---|
| Input, $ per 1M tokens | 0.10 (up to 100K) / 0.50 (above) | 1.00 | 2.00 |
| Output, $ per 1M tokens | 0.50 / 2.50 | 5.00 | 10.00 |
| Cache read, $ per 1M | 0.01 / 0.05 | 0.10 | 0.10 |
| Cache write, $ per 1M | 0.125 / 0.625 | 1.25 | 2.50 |
| Context | 1M tokens | 200,000 tokens | — |
| Maximum response | 128,000 tokens | — | — |
| Effort setting | Low, Med, High, Xhigh, Max | no | yes |
Sources: prices and effort — Claude Haiku 5.5 page; context and response length for Haiku 5.5 — DataNorth; Haiku 4.5 context — ComputingForGeeks. A dash means the parameter was not specified in the sources used. According to DataNorth, batch processing gives an additional 50% discount.
Two more technical details that matter during implementation, according to DataNorth. The temperature, top_p and top_k parameters should be left at their default values: if you try to change them, the API returns an error. And the model deprecation date is set no earlier than October 7, 2027 — this is a planning reference, not a guarantee of longer support.
The 100,000-token threshold: where a cheap model gets expensive
The 100,000-token threshold is the length limit for a single request, beyond which Haiku 5.5 is billed at five times the normal rate. According to ComputingForGeeks, the higher rate applies to all tokens in that request, including the response and cache, not just the overage. This is the only current Claude model with tiered pricing by context length.
In practice, that means three things:
- A 1-million-token context is not a reason to fill it. The model will accept a long contract or an exported chat log in full, but you will pay the higher rate for it.
- Long documents are cheaper when split up. If the task is to find account details or liability clauses in a document, splitting it into chunks under 100,000 tokens is usually cheaper than sending one large request. You still need to verify quality separately.
- Context compression in agents is a must-have setting. An agent that accumulates chat history will eventually cross the threshold. Compressing history is one of the use cases Anthropic positions Haiku for.
According to VentureBeat, about 90% of Haiku 4.5 requests fell into the short category. For high-volume routine work, the threshold is unlikely to be a problem, but for long-document workflows it should be considered from day one.
How much a task costs: calculation examples
Below are our examples based on published rates. Token volumes are hypothetical: real numbers depend on language, format, and the tokenizer, so they should be measured on your own data. We wrote about why Russian text usually takes more tokens than English in our article on Russian text tokenization.
Example 1. Support ticket classification. 10,000 tickets per month; for each one, 2,000 input tokens (instructions plus email text) and 300 output tokens (category, priority, short summary).
| Model | Monthly input | Monthly output | Total per month |
|---|---|---|---|
| Haiku 5.5 | 20M × $0.10 = $2.00 | 3M × $0.50 = $1.50 | $3.50 |
| Haiku 4.5 | 20M × $1.00 = $20.00 | 3M × $5.00 = $15.00 | $35.00 |
| Sonnet 5.5 | 20M × $2.00 = $40.00 | 3M × $10.00 = $30.00 | $70.00 |
If the shared instruction takes up 1,500 of the 2,000 input tokens and is cached, the input cost on Haiku 5.5 drops even further: 15 million tokens from cache would cost $0.15 instead of $1.50 (not including the one-time cache write).
Example 2. One long document. Let's compare a request with 99,000 input tokens to one with 120,000; in both cases, the response is 2,000 tokens.
| Request | Input | Output | Total per request |
|---|---|---|---|
| 99,000 tokens (below the threshold) | 99,000 × $0.10 / 1M = $0.0099 | 2,000 × $0.50 / 1M = $0.0010 | $0.0109 |
| 120,000 tokens (above the threshold) | 120,000 × $0.50 / 1M = $0.0600 | 2,000 × $2.50 / 1M = $0.0050 | $0.0650 |
A request that is only 21% longer costs almost six times more. For a single document, that's pennies; for a pipeline handling tens of thousands of documents, it's already a budget item.
The takeaway from the calculations: for short, high-volume tasks, the API cost of Haiku 5.5 is so low that the project's main expenses will be elsewhere — in system integration, data preparation, quality checks, and support. More on that in our article about the cost of AI implementation.
Test results and what they mean
All figures below were published by Anthropic; we could not find independent verification as of the publication date. The comparison with GPT-6 Luna also comes from Anthropic, and, as DataNorth notes, the company does not specify whether it ran the competitor itself.
| Test | What it measures | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|---|
| GDPval-AA v2.1 | Office work, Elo score | 1620 | 735 | 1437 | 1840 |
| AA-Briefcase v1.1 | Office work | 1578 | 614 | 1336 | 1824 |
| OSWorld 2.1, offline | Computer control | 72.4% | 15.7% | 48.9% | 83.9% |
| Humanity's Last Exam, no tools | Hard questions | 45.9% | 10.2% | — | 56.9% |
| Terminal-Bench 4.0 | Agentic coding | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 | Coding | 46.4% | — | 42.4% | 52.1% (Xhigh) |
How businesses should read this:
- Office tasks and interface work — the biggest gain. This directly applies to document processing, form filling, and browser actions.
- Agentic coding remains a weak spot compared with Sonnet 5.5. For code changes in a repository, the smaller model is better suited as an assistant than as the primary executor.
- Anthropic customer feedback sounds impressive: HubSpot reported 92.8% on average across three runs of its CRM task set, and AlphaSense reported 0.84 versus 0.76 for Haiku 4.5 on 400 queries. But these are proprietary company benchmarks with unknown methodology; you cannot assume the same results in your own workflow.
Reasoning depth modes
Effort is a setting for how much the model “thinks” before answering. The higher the level, the better the result on difficult tasks, and the more tokens and time the response takes. Haiku 5.5 has five levels: Low, Med, High, Xhigh, and Max; the default is Medium.
The difference can be significant. According to VentureBeat, the result in Terminal-Bench 4.0 is about 39% at the maximum level and about 20% at the medium level. Vendor tables often show the best mode, while in production the model runs at the default setting.
A practical rule of thumb: for classification and field extraction, start with Low or Med and raise the level only if testing on your own examples shows errors. For multi-step reasoning tasks, compare Med and High for quality and cost on the same dataset.
Which business tasks Haiku 5.5 is good for
The model is cost-effective where there are many tasks, each one is short, and the correct answer can be checked against clear rules.
| Task | Why it fits | What to check |
|---|---|---|
| Sorting inquiries and emails | Short input, answer from a fixed list of categories | Accuracy on borderline categories |
| Extracting fields from documents | Invoices, acts, applications: a known set of fields | Missing and invented values |
| Short summaries of calls and correspondence | High-volume operation, low cost of mistakes | Whether agreements and deadlines are not lost |
| Routing requests between models | Fast first step, low-cost complexity assessment | Share of tasks incorrectly “simplified” |
| Support chat responses | Speed matters more than depth | Working with the knowledge base and escalating to a human |
| Sub-agent in an agent system | Search, retrieval, context compression | Quality of result transfer to the main model |
| Browser actions | Strong Gains in OSWorld | Access Permissions and Action Confirmation |
Many of these scenarios are covered in detail in our roundup of use cases for small language models: the same principles apply to Haiku 5.5 and to open models that you can deploy on your own infrastructure.
The Planner-and-Executor Pattern
The planner-and-executor pattern is an agent architecture in which a powerful model breaks a task into steps, while a low-cost, fast model carries out many small steps. Anthropic explicitly recommends Haiku 5.5 as a subagent with Opus 5.5 and Sonnet 5.5. In examples on the model page, Rogo uses it to extract data from annual reports, and Cognition uses it as an assistant in its Devin agent.
What this looks like in a business workflow using incoming request processing as an example:
- Haiku 5.5 reads each email, identifies the request type, and extracts the key details.
- Simple requests are sent straight to the CRM or back-office system using a template.
- Complex or nonstandard requests with the collected data are passed to Sonnet 5.5 or a human.
- The strong model drafts the response or action plan, and Haiku 5.5 again handles routine steps: searching the database, filling in fields, and summarizing the history.
We choose this setup instead of a single powerful model for everything because most steps in a typical process are simple, and paying premium rates for them does not make sense. The trade-off is complexity: you add a routing rule that can also make mistakes, and it needs to be monitored. We covered how to set up routing and what events to log in our article on AI model routers.
When Haiku 5.5 Is Not the Right Choice
- Complex agent code. The gap with Sonnet 5.5 on Terminal-Bench is nearly twofold, and Anthropic itself recommends higher-tier models for these tasks.
- Penetration testing and offensive techniques. According to Anthropic, Haiku 5.5 has stricter cybersecurity safeguards than Haiku 4.5: the model allows more defensive tasks but blocks pentesting. Expanded access is available only through Anthropic's verification programs.
- Ongoing work with prompts longer than 100,000 tokens. Here it is worth comparing it with GPT-6 Luna: its base rates are the same, $0.10 and $0.50, and the higher tier starts after 272,000 tokens, according to VentureBeat. Our breakdown of the lineup is in the article on GPT-6 Sol and Luna.
- High-stakes decisions without review. Medical, legal, and financial conclusions from a smaller model should be double-checked by a human or a stronger model.
- Data that cannot be transferred abroad. For personal data and trade secrets, the issue is not model quality but where the data is processed.
Moving from Haiku 4.5: what to check
If Haiku 4.5 is already running in your workflows, the move may look like a simple ID swap, but there are a few things worth checking first:
- Token counting. Anthropic says the new tokenizer uses slightly more tokens for the same task. DataNorth estimates the difference at about 30% on the same text. Measure it on your own data before recalculating the budget.
- Generation parameters. If the code passes
temperature,top_portop_k, the requests will start returning an error. Remove these parameters. - Effort level. Set it explicitly for each task instead of relying on the default value.
- Prompt length. Find prompts longer than 100,000 tokens in the logs: for those, the savings will be 50%, not 90%.
- Response format. Run the test set and compare the JSON structure, response length, and failure rate: the new model may respond differently to the same instructions.
- Failure behavior. If your tasks involve information security, check whether they fall under stricter restrictions.
Availability in Russia and data requirements
Anthropic does not include Russia in its list of supported regions, and in early October 2026, according to user reports, the company carried out another wave of account bans for accounts operating through VPNs. We explained in detail what this means for companies and how to avoid disrupting workflows in our article on Claude being blocked in Russia.
For business, this leads to three takeaways:
- Bypassing blocks is not a foundation for a production workflow. Access can disappear at any moment, along with history and settings.
- Transferring personal data to an overseas service requires a separate assessment. Details on cross-border data transfer conditions are in the article on Federal Law 152-FZ and foreign AI. This material is not legal advice.
- The planner-and-executor pattern can be built with available models. The role of Haiku in it can be filled by open models deployed on your own servers or in a Russian cloud; the options are described in the article on running open LLM inference in a Russian cloud.
For a Russian company, Haiku 5.5 is useful as a benchmark: it shows what quality and what price are currently available for high-volume tasks and sets the standard for choosing a replacement.
How to Test a Model on Your Own Tasks
Public benchmarks answer the question of which model is stronger on average, but not which one will fit my workflow. Here is the testing process we recommend:
- Choose one process with a high volume of repetitive tasks, such as sorting incoming emails.
- Build a validation set of 100 to 300 real examples with correct answers, including edge cases and rare cases. Anonymize personal data.
- Set acceptance criteria in advance: an acceptable error rate, unacceptable errors (for example, missing an urgent complaint), maximum response time, and cost per 1,000 tasks.
- Run the candidates — Haiku 5.5 or the replacement available to you, your current model, and one open model — on the same set with the same instructions. For Haiku, compare at least two effort levels.
- Calculate cost based on actual tokens, not estimates, and separately track the share of prompts longer than 100,000 tokens.
- Review the errors manually. It matters not only how many errors there are, but what kind: a fabricated value in a contract field is more dangerous than the wrong email category.
Limitation of the method: a validation set of a few hundred examples will not reveal rare failures. That is why you need a log of model decisions after launch and spot checks by a human.
Frequently Asked Questions
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for requests up to 100,000 tokens. For longer requests, $0.50 and $2.50. Cache reads cost $0.01 per million tokens, and batch processing, according to DataNorth, comes with a 50% discount.
How is Haiku 5.5 better than Haiku 4.5?
According to Anthropic, the model is on average about 75% cheaper to run and significantly stronger: the GDPval-AA score rose from 735 to 1620, and the OSWorld 2.1 result went from 15.7% to 72.4%. It also adds up to 1 million tokens of context and a reasoning depth setting.
Which should I choose: Haiku 5.5 or Sonnet 5.5?
Haiku 5.5 is for high-volume short tasks with verifiable answers: classification, field extraction, summaries, and subagent roles. Sonnet 5.5 is for complex multi-step tasks and agentic coding, where, according to Anthropic, it is almost twice as strong on Terminal-Bench 4.0.
How is Haiku 5.5 different from GPT-6 Luna?
The base prices are the same — $0.10 and $0.50 per million tokens. In tests published by Anthropic, Haiku 5.5 outperforms GPT-6 Luna in office work and computer control. But GPT-6 Luna’s higher rate starts after 272,000 tokens, while Haiku 5.5’s starts after 100,000.
Can Claude Haiku 5.5 be used in Russia?
Officially, Anthropic does not support Russia, and accounts operating through VPNs are blocked. For business workflows, it is better to choose a legally available model or deploy an open model on your own servers, and use Haiku 5.5 as a benchmark for quality and price.
What context window does Claude Haiku 5.5 have?
Up to 1 million input tokens and up to 128,000 tokens in the response, according to DataNorth. But requests longer than 100,000 tokens are billed five times higher, so it makes sense to use the long context only where it is truly necessary.
Primary sources and verification method
We read the primary sources on October 11, 2026 and cross-checked prices, specifications, and test results across them. Prices, effort levels, test results, and task recommendations were taken from Anthropic’s model page; response length, context, API limits, and the deprecation date were taken from recaps by industry publications, because these details are not listed on the model page. The cost calculations are our own examples based on the published rates; we did not conduct our own model testing.
- Anthropic: Claude Haiku 5.5 — prices, tests, tasks, safety limits, availability.
- VentureBeat, 07.10.2026: Anthropic launches Claude Haiku 5.5 — share of short requests, price comparison, results at different effort levels.
- DataNorth, 08.10.2026: Anthropic releases Claude Haiku 5.5 — context, response length, parameter limits, deprecation date.
- ComputingForGeeks: Claude Haiku 5.5 released — rule for applying the higher rate, Haiku 4.5 context window.
Limitations: the tests were published by the vendor and have not been independently verified; cloud platform pricing and contract rates may differ from Anthropic’s list price; the blocking data is based on user reports and media coverage.
How AIrassvet helps you choose a model for high-volume tasks
If you want to use AI to sort incoming requests, extract data from documents, or summarize conversations, model selection is only part of the work. On the AIrassvet website the following services are described that relate directly to this task:
- Deployment and fine-tuning of language models, including on company servers, when data cannot be sent to an external service.
- Development of AI agents integrated with communication channels and business systems to process requests, documents, and correspondence.
- Process audit and data preparation before launch.
- Testing, launch, and post-deployment support.
The first step is to choose one process, describe its current metrics, data sources, constraints, and the criterion by which you will consider the model’s result acceptable.
Bottom line: a low-cost model for volume, not for everything
Claude Haiku 5.5 has made large-scale text processing almost free from an API cost standpoint: for short tasks, it is ten times cheaper than Haiku 4.5 and twenty times cheaper than Sonnet 5.5 on input tokens, and the published tests show it is noticeably stronger than its predecessor. At the same time, it has two limitations that matter more than the marketing numbers: the rate is five times higher after 100,000 tokens, and it lags significantly behind larger models on complex agentic tasks.
For a Russian company, the practical takeaway is clear: the setup where a low-cost model handles routine work and a stronger model handles complex steps should be built now, using legally available models. Start with one high-volume process and a test set — that will show which model works and how much 1,000 tasks cost.
