Claude 3.5 Haiku

What it is

Claude 3.5 Haiku is a small, fast language model from Anthropic. It belongs to the Haiku tier of the Claude family, the tier used for the company's quickest and cheapest models, as distinct from the larger Sonnet and Opus tiers. Anthropic called it the next generation of its fastest model, a direct successor to Claude 3 Haiku^[1].

It was announced on 22 October 2024, in the same release that introduced an upgraded Claude 3.5 Sonnet and computer use^[1].

Availability

The model became available on 4 November 2024 through the Anthropic API, Amazon Bedrock and Google Cloud Vertex AI^[1][2]. At that point it accepted text only, with image input announced to follow^[1][2]. Its training data ran to July 2024^[5].

What Anthropic emphasised

Anthropic said the model improved over Claude 3 Haiku across every skill set, and that it was particularly strong on coding tasks^[1]. It also said it follows instructions more strongly and uses tools more accurately^[1]. Tool use means a model can call external functions or services to get work done, rather than only producing text.

On SWE-bench Verified, a benchmark that asks a model to resolve real software issues taken from code repositories, Claude 3.5 Haiku scored 40.6%^[1][4]. Anthropic said that score outperforms many agents built on publicly available state-of-the-art models, including the original Claude 3.5 Sonnet and GPT-4o^[1]. An agent here means a system that lets a model work in several steps, reading files, running commands and checking results, instead of answering in one pass. The comparison is a way of saying that a small model inside a good harness can beat a larger model used less carefully.

Anthropic further said the model matched or surpassed Claude 3 Opus, the largest model of the previous generation, on many intelligence benchmarks, while running at a speed similar to Claude 3 Haiku^[1][4]. Small and fast, yet at the level of the previous flagship: that is the claim the release was built around, and the reason a model like this matters, since most real applications care as much about latency and cost as about raw capability.

Pricing

At availability, Anthropic priced Claude 3.5 Haiku at $1 per million input tokens and $5 per million output tokens^[2]. That was four times the price of Claude 3 Haiku, and the company cited the model's higher intelligence as the reason^[2]. On 3 December 2024 the price was revised to $0.80 per million input tokens and $4 per million output tokens^[1].

Where it sits in the lineup

Claude 3.5 Haiku belongs to the 3.5 generation of Claude models, which also included the upgraded Claude 3.5 Sonnet. Its job in the lineup was to be the cheap, quick option: the model a developer reaches for when a task has to be done many times, quickly and at low cost, rather than the one used for the hardest single request. Comparing speed with Claude 3 Haiku and intelligence with Claude 3 Opus frames it in exactly those terms.

The model was succeeded by Claude Haiku 4.5 on 15 October 2025^[5]. Deprecation was announced on 19 December 2025, and the version identifier claude-3-5-haiku-20241022 was retired on 19 February 2026^[3].