Claude 3 Haiku becomes available

Background

Claude 3 Haiku was the Haiku model of the Claude 3 line, a model built around speed and cost.^[1]

What happened

On 13 March 2024 Anthropic made Claude 3 Haiku available. The company called it the fastest and most affordable model in its intelligence class.^[1]

It launched in the Claude API, for Claude Pro subscribers on claude.ai, and on Amazon Bedrock. Google Cloud Vertex AI was to follow.^[1]

Key details

Haiku processes about 21K tokens, around 30 pages, per second for prompts under 32K tokens. A token is a small piece of text, roughly a word or part of a word.^[1]

Its pricing uses a one-to-five ratio between input and output tokens. Anthropic said that analysing 400 Supreme Court cases, or 2,500 images, would cost about $1.^[1]

The model supports image input.^[1]

Anthropic also highlighted enterprise security measures for the model. These included continuous monitoring, endpoint hardening, encryption and penetration testing.^[1]

Impact

Anthropic presented Haiku on the basis of speed and price. A ratio of one to five between input and output prices, together with the $1 example, showed the cost of large-scale analysis at a glance.^[1]

Tokens per second is a measure of how quickly a model processes text. A model that handles more tokens per second can read long material and return results sooner, which matters for tasks that process many documents.

The one-to-five ratio meant output was priced at five times the rate of input. Because answers are output, shorter answers cost less, and a task with a short reply and a long document costs mostly on the input side.

Penetration testing means hiring testers to attack a system in order to find weaknesses before others do.

Tiers matter when choosing a model. A smaller model can be the right choice when a task is repetitive and speed or cost matters more than the deepest reasoning.