# What is Claude Haiku 4.5, and when is a small, fast model the right choice?

Date: 2025-10-16 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-claude-haiku-4-5/

**Answer:** Claude Haiku 4.5 is Anthropic's new small model, released on October 15, 2025, which Anthropic says codes about as well as Claude Sonnet 4 did five months ago, while costing a third as much and running over twice as fast. It is available to all Claude users, including the free plan, and costs developers $1 per million input tokens and $5 per million output tokens.

Yesterday, Wednesday, October 15, 2025, Anthropic released **Claude Haiku 4.5**, a fresh release in its smallest model family. Anthropic sells three sizes of Claude: Haiku is the small, quick one, Sonnet the middle and Opus the largest. Small models matter because many real tasks need answers that are fast and cheap more than answers that are as deep as possible.

Anthropic's summary is that capability which counted as cutting-edge a few months ago is now cheaper and quicker. Five months ago, Claude Sonnet 4 was among the best coding models available; Haiku 4.5 comes close to that coding level while costing a third as much and running over twice as fast. Computerworld notes that the previous Haiku came out a year ago.

## What Anthropic released

| | Claude Haiku 4.5 |
|---|---|
| Model name for developers | `claude-haiku-4-5` |
| Price per million tokens (input / output) | $1 / $5 |
| SWE-bench Verified (fixing real code bugs) | 73.3%, Anthropic's figure |
| Where | Claude apps (all plans, including free), Claude Code, Claude API, Amazon Bedrock, Google Cloud Vertex AI |
| Safety level | ASL-2 |

A **token** is a small chunk of text, roughly part of a word. **SWE-bench Verified** asks a model to fix real bugs from open-source projects and checks the fixes with tests.

Anthropic adds that Haiku 4.5 even beats Sonnet 4 at some tasks, such as **computer use**, where the model operates a computer by looking at the screen and clicking and typing. That makes tools like the Claude for Chrome browser extension quicker. According to Computerworld, several benchmark results put it level with Sonnet 4, OpenAI's GPT-5 and Google's Gemini 2.5, including SWE-bench and Terminal-Bench, a test of work in the command line.

Anthropic's most capable model remains Claude Sonnet 4.5, which arrived two weeks earlier. Haiku 4.5 is offered as the choice when you want near-top performance at a much lower price.

## Why speed and cost matter

Every reply from a model has a **latency**, the wait before the answer arrives, and a price based on tokens. For a customer-service chat, a coding assistant that suggests edits as you type, or a voice helper, a reply that arrives a second sooner feels very different. For a business running millions of requests, a model that costs a third as much changes what is affordable.

A quick calculation shows the scale. Suppose an app sends Haiku 4.5 one million tokens of input and receives 200,000 tokens of output in a day:

- Input: 1 million × $1 per million = $1.00
- Output: 0.2 million × $5 per million = $1.00
- Daily total: **$2.00**

Customers quoted by Anthropic tell the same story. Gamma, which makes presentations, says Haiku 4.5 followed instructions for slide text with 65% accuracy against 44% for its premium model, and GitHub says it brings Copilot quality comparable to Sonnet 4 at a faster speed.

## A worked example: one planner, many helpers

The most interesting idea in the announcement is using models together. A larger model can act as a planner while several small models do the pieces at once. These helpers are often called **sub-agents**: agents launched by another agent to handle part of a job. Anthropic's example is Claude Sonnet 4.5 breaking a complex problem into a multi-step plan, then directing a team of Haiku 4.5 models to complete the parts in parallel.

Here is how that could look for updating a website's documentation:

| Role | Model | Job |
|---|---|---|
| Planner | Claude Sonnet 4.5 | Read the whole codebase, list the 12 pages that need updates, and write instructions for each |
| Helpers 1–12 | Claude Haiku 4.5 | Each rewrites one page following its instructions, at the same time |
| Reviewer | Claude Sonnet 4.5 | Check the 12 drafts for consistency and send any weak ones back |

The expensive model does the thinking that needs depth; the cheap, fast model does the repeated work. Computerworld makes the same point: because Haiku uses few resources, several copies can run in parallel or alongside larger models.

> **Try it:** Take a task you do in many small, similar pieces, such as writing product descriptions or summarizing meeting notes. Write one short instruction for a single piece, run it with Claude Haiku 4.5 on five examples, and check how many need fixing.

## Safety

Anthropic ran its usual safety and alignment tests and reports low rates of worrying behavior. On its automated alignment measure, Haiku 4.5 showed fewer misaligned behaviors than both Sonnet 4.5 and Opus 4.1, which Anthropic says makes it its safest model by that metric.

Because testing found only limited risk of helping with chemical, biological, radiological or nuclear weapons, Anthropic released it under **AI Safety Level 2** (ASL-2), a lighter set of protections than the ASL-3 standard that covers Opus 4.1 and Sonnet 4.5. The details are in the model's system card.

## When should you choose Claude Haiku 4.5 over a larger model?

Choose Haiku 4.5 when speed and volume matter more than maximum depth: chat assistants, customer support, quick coding help, sorting or summarizing large numbers of items, and helper agents inside a larger workflow. If you are on the free Claude plan, it is now a capable everyday model; in Claude Code, it makes back-and-forth work feel more responsive and lets you do more within your usage limits.

Choose Sonnet 4.5 or Opus when a task is long, unfamiliar or high-stakes, such as a big change across many files, careful research or a decision you cannot easily check. A good habit is to try Haiku 4.5 first on a sample of your real work. If its answers hold up, you gain speed and save money; if not, move the hard parts to a larger model and keep Haiku for the rest.

## Sources

- Introducing Claude Haiku 4.5: https://www.anthropic.com/news/claude-haiku-4-5
- Anthropic releases new version of its smaller Haiku model: https://www.computerworld.com/article/4073957/anthropic-releases-new-version-of-its-smaller-haiku-model.html