Claude Haiku 5.5: The Small Model Built for Big Workloads

Picture a support queue filling up faster than a human team can read it. Each request may be modest—a short summary, a category, a database lookup—but thousands of those requests turn model choice into an engineering and budgeting problem. A model that saves a fraction of a second on every call can change the feel of an entire product.

Anthropic introduced Claude Haiku 5.5 on October 7, 2026, as its fastest, least expensive, and most capable small model so far. The interesting story is not that it replaces larger models. It gives developers a practical first layer for repetitive work, quick agent turns, and tasks where response time matters. (anthropic.com)

What a small model means here

A large language model, or LLM, is a system trained to generate and understand text, code, images, or other data. Small model describes Claude Haiku 5.5’s position within Anthropic’s model family, not a lack of useful capability. It is designed to handle narrow jobs repeatedly, while larger models such as Sonnet and Opus take on difficult planning, ambiguous judgment, and long-running work.

That division matters in an AI agent. An agent is software that calls a model, uses tools, reads data, and takes several steps toward a goal. A subagent is a helper model call given one bounded piece of that work. A larger model might plan a coding task, then send Haiku 5.5 to inspect a file, summarize a log, or extract one value from a long report.

Haiku 5.5 also fits well with compaction. Compaction means reducing an older conversation or work history into a shorter summary before an agent continues. The summary does not need grand reasoning; it needs to preserve the decisions, open tasks, and important facts accurately and quickly.

The headline improvement is cost per useful result

A token is a small piece of text that a model processes and that an API provider uses for billing. Claude Haiku 5.5’s published prices are especially low for prompts up to 100,000 tokens:

Prompt caching means saving a repeated part of a request—such as a long system instruction or shared document—so later calls can reuse it at a lower price. That can matter in applications where every request includes the same policy manual, product catalog, or tool description.

Anthropic says Haiku 5.5 costs around 75% less to run than Haiku 4.5 on average. The detail behind that number matters: list prices are 90% lower for requests up to 100,000 tokens and 50% lower above that threshold, while changes to the tokenizer affect how many tokens a task uses. In other words, compare complete task costs rather than multiplying a headline price by an old token count.

For a rough example, imagine a batch of requests whose prompts are each under 100,000 tokens. One million input tokens would cost $0.10, and 200,000 output tokens would cost another $0.10, for $0.20 before any cache charges. At high volume, that difference can decide whether a feature runs continuously or only when a customer pays for it.

Adjustable effort turns one model into a dial

Haiku 5.5 is the first Haiku-class model with an adjustable effort setting. Effort is a control over how much the model optimizes for intelligence versus cost and speed. Lower effort suits predictable classification or short extraction; higher effort gives the model more room for a difficult decision with ambiguous evidence.

That makes the model more useful than a single fixed speed setting. A production system can use lower effort for routine tickets and raise it for a support request that contains several possible causes. The right choice should come from an evaluation set—a collection of representative examples kept separate from the prompts used during development—not from intuition alone.

The basic Messages API pattern looks like this:

import anthropic

client = anthropic.Anthropic
ticket = 'The mobile app logs me out after every update.'

message = client.messages.create(
 model='claude-haiku-5-5',
 max_tokens=120,
 system=(
 'Classify the support ticket. Return JSON with '
 'category, urgency, and one-sentence summary.'
 ),
 messages=[
 {
 'role': 'user',
 'content': f'<ticket>{ticket}</ticket>',
 }
 ],
)

print(message.content[0].text)

The system prompt contains the application’s standing instructions, while the user message carries the individual ticket. The XML-style tag marks the input boundary so the model can distinguish the ticket from the instructions. Anthropic’s current SDK examples use this same general message structure, with a model name, output limit, system instructions, and user content. (docs.anthropic.com)

Benchmarks show a jump, with a clear boundary

Anthropic’s published results show a large improvement over Haiku 4.5. On GDPval-AA v2.1, a test of professional knowledge work, Haiku 5.5 scored 1,620 compared with 735 for Haiku 4.5. On the offline subset of OSWorld, which measures computer use across long, multi-step tasks, it scored 72.4% versus 15.7%. On Terminal-Bench 4.0, an evaluation of command-line agent work, it reached 39.2%, while Haiku 4.5 scored 0.0%. (anthropic.com)

Those numbers are encouraging, but they do not mean Haiku 5.5 is the best choice for every agent. Anthropic reports Sonnet 5.5 at 70.6% on Terminal-Bench, well above Haiku 5.5. The practical lesson is routing: use Haiku for bounded subtasks, and move complex, multi-step coding or judgment calls to a larger model. Benchmarks are useful instruments, but your own workload remains the road test.

A sensible place for Haiku 5.5 in production

A dependable architecture often looks like a relay. Sonnet or Opus creates a plan. Haiku 5.5 handles the many small operations inside that plan. A larger model or a deterministic validator checks the final result when an error would be expensive.

Good candidates include:

• classifying incoming requests;
• summarizing documents or conversation history;
• extracting a field from a known report;
• generating short database-query explanations;
• handling fast customer-support replies;
• operating a browser or computer for narrowly scoped tasks.

For each workflow, measure more than latency, which is the time a request takes to return. Track task success, false positives, false negatives, tool errors, and cost per successful outcome. A cheap model that sends the wrong refund policy to a customer is not cheap in practice.

Capability also does not remove the need for safeguards. Anthropic reports improved alignment evaluations for Haiku 5.5 and says its cybersecurity protections are stricter than Haiku 4.5’s, while still restricting penetration testing and other high-risk techniques. Applications should add their own permission checks, structured-output validation, tool allowlists, and human review for irreversible actions.

Claude Haiku 5.5 is available through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Its strongest role is not as a cheaper copy of a frontier model. It is the fast worker that makes thousands of small, useful decisions economically practical.