Claude 3.5 Haiku
Anthropic's fast, low-latency model for coding, tools and user-facing products
Claude 3.5 Haiku is a small, quick language model from Anthropic, announced in October 2024 and released later that month. The company says it reaches the level of its earlier largest model, Claude 3 Opus, on many evaluations while keeping the speed of the Haiku line, and it is aimed at coding, tool use and personalised experiences over large data sets.
- Announced in October 2024 as the successor to Claude 3 Haiku
- Anthropic claims it matches or beats Claude 3 Opus on many evaluations
- Scores 40.6% on SWE-bench Verified, per the company
- Priced at $0.80 per million input tokens and $4 per million output tokens from 3 December 2024
- Text-only at launch, with image input planned
What it is
Claude 3.5 Haiku is the speed-focused member of Anthropic's Claude 3.5 generation. Announced alongside an upgraded Claude 3.5 Sonnet and a computer use beta, it was due to appear later in the same month. Anthropic positions it as the successor to Claude 3 Haiku, running at roughly the same pace while improving across every skill area. The company claims it beats Claude 3 Opus, the biggest model of the previous generation, on many intelligence evaluations. It is aimed at customer-facing products, specialised sub-agents and turning large record sets such as purchase history, pricing and inventory into personalised output. Coding is a stated strength.
How it came about
Anthropic announced Claude 3.5 Haiku in October 2024 in the same post as the refreshed Claude 3.5 Sonnet and the computer use public beta. The model was scheduled for release later that month through the company's own API, Amazon Bedrock and Google Cloud's Vertex AI. It began as a text-only offering, with image input promised afterwards. On 3 December 2024 the company revised its pricing, setting the model at $0.80 per million tokens of input and $4 per million tokens of output.
Inside the model
Anthropic describes Claude 3.5 Haiku as the next step in its fastest model line, keeping latency low while raising instruction following and the accuracy of tool calls. The company says it reaches or passes Claude 3 Opus on many intelligence benchmarks at a speed close to Claude 3 Haiku. It ships first for text, with picture input to follow. No layer counts, parameter totals, training corpus or compute figures were published for it. The same announcement introduced a computer use interface that lets Claude read a screen, move a cursor, click and type, though that capability was presented with the Sonnet model.
What it can do
Anthropic reports a score of 40.6% on SWE-bench Verified, a test of fixing real software issues, and says this beats many agents built on widely available leading models, naming the first Claude 3.5 Sonnet and GPT-4o among them. The company claims gains over Claude 3 Opus on many intelligence evaluations while matching the earlier Haiku's speed. It highlights coding, tool use and instruction following, and suggests the model suits user-facing products, sub-agent roles and personalisation over large data sets.
How to use it
The model was to reach Anthropic's first-party API, Amazon Bedrock and Google Cloud's Vertex AI later in the month of announcement, starting as text-only with image input to come. After a pricing revision on 3 December 2024, the company set rates at $0.80 per million input tokens and $4 per million output tokens. No downloadable weights or open licence were offered.
Safety and testing
Anthropic says joint pre-deployment testing of the upgraded Claude 3.5 Sonnet was carried out with the US AI Safety Institute and the UK Safety Institute, and that the ASL-2 standard from its Responsible Scaling Policy still applied to that model. For the computer use capability shown alongside Haiku, the company built classifiers to detect when the feature is in use and whether harm is occurring, and it warned that actions such as scrolling, dragging and zooming remain difficult, advising developers to start with low-risk tasks.
Reception and debate
The announcement quotes early users of the upgraded Claude 3.5 Sonnet rather than Haiku. GitLab reported stronger reasoning, by up to 10% across use cases, with no added latency in DevSecOps work. Cognition saw sizeable gains in coding, planning and problem solving. The Browser Company said the model beat everything it had tested for automating web workflows. No separate reception for Claude 3.5 Haiku appears in the available material.
Why it matters
Anthropic presents Claude 3.5 Haiku as bringing the level of its former largest model into a fast, low-cost tier, which it says suits customer-facing products, specialised sub-agents and personalisation drawn from large data sets such as purchase history, pricing and inventory. It arrived in the same release as the computer use public beta, which the company describes as early and error-prone but intended to improve with developer feedback.
Questions people ask about Claude 3.5 Haiku
What is Claude 3.5 Haiku?
It is the quick, low-latency model in Anthropic's Claude 3.5 generation, announced in October 2024. The company says it improves on Claude 3 Haiku across every skill area while keeping similar speed, and reaches the level of Claude 3 Opus on many intelligence evaluations.
How much does it cost?
Anthropic revised the rates on 3 December 2024 to $0.80 per million tokens of input and $4 per million tokens of output. Earlier figures in the announcement were superseded by that change.
What can it do?
Anthropic highlights coding, tool use and instruction following. It reports 40.6% on SWE-bench Verified and suggests the model for user-facing products, specialised sub-agents and personalised output built from large record sets such as purchase history, pricing and inventory.
How do you get it?
It was planned for Anthropic's own API, Amazon Bedrock and Google Cloud's Vertex AI later in the month of announcement. It launched as text-only, with image input to follow. No open weights were released.
Does it use computers like a person?
Computer use was introduced in the same announcement as a public beta on the API, letting Claude view a screen, move a cursor, click and type. Anthropic presented it with Claude 3.5 Sonnet and called it experimental, cumbersome and error-prone at times.
Good reading and sources
There is no English Wikipedia article on Claude 3.5 Haiku yet, so the facts on this page come from the sources listed above: anthropic.com/news/3-5-models-and-computer-use. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.