Claude 3 Haiku

What Claude 3 Haiku is

Claude 3 Haiku is a large language model from Anthropic and the smallest and fastest member of the Claude 3 family. Anthropic announced the family on 4 March 2024, and Haiku became available on 13 March 2024, released alongside Claude 3 Sonnet and Claude 3 Opus.^[1][2]

The model accepts image input as well as text, and works within a 200,000-token context window.^[1] A context window is the amount of material a model can hold in view at once; 200,000 tokens is enough for long documents, code bases or a batch of images to be supplied in a single request. Anthropic describes Haiku as the fastest and most affordable model in its intelligence class, and prices it at half the cost of other models in its performance tier.^[2]

Speed

Anthropic's pitch for Haiku centred on throughput. According to the company, it processes about 21,000 tokens — roughly 30 pages of text — per second on prompts under 32,000 tokens, which Anthropic says makes it about three times faster than its peers.^[2] A token is the unit a language model reads and writes in, typically a fragment of a word, so a high token rate translates into short waiting times for the user.

The company also gave a reading example: the model can work through an information-dense research paper of about 10,000 tokens, charts included, in under three seconds.^[1] Latency of that order matters for uses where the model sits inside another piece of software — a chat interface, a search box, a coding assistant — and every extra second is felt by the person waiting.

Price

Pricing was set at $0.25 per million input tokens and $1.25 per million output tokens, a ratio of one to five between what goes in and what comes out.^[1][2] Because output tokens cost five times as much as input tokens, the cost of a workload depends heavily on how much text the model is asked to generate rather than read.

Anthropic illustrated the scale with two examples: analysing 400 Supreme Court cases, or 2,500 images, would cost about $1.^[2] Figures like these are what made the small-model tier interesting for bulk processing — classification, extraction and summarisation over large collections, where the bill is dominated by volume rather than by the difficulty of any single request.

Availability and training

At launch the model was available in the Claude API and to Claude Pro subscribers on claude.ai, and on Amazon Bedrock, with support on Google Cloud Vertex AI to follow.^[2] That distribution put it in front of both developers building on Anthropic's own interface and teams already working inside Amazon's cloud platform.

According to Wikipedia, its training data ran to August 2023.^[4] Anthropic said the model underwent rigorous testing to reduce harmful outputs and jailbreaks.^[2]

Place in the line and retirement

Haiku sat at the bottom of the Claude 3 range, the tier sold on speed and cost while Sonnet and Opus occupied the larger end. The name carried forward into later generations: when Anthropic announced the model's deprecation on 19 February 2026, it recommended Claude Haiku 4.5 as the replacement. The company retired the model identifier claude-3-haiku-20240307 on 20 April 2026.^[3]

The gap between announcement and retirement — roughly two years — is a reminder that model names in an API are versioned artefacts: applications that call a specific identifier have to be moved to a successor when that identifier is switched off.