The Claude 3 Model Family (2024)

Anthropic's Claude 3 technical report describes a family of three multimodal models (Haiku, Sonnet, Opus) that set new standards for safety, honesty, and capability, with particular emphasis on reducing both harmful outputs and unnecessary refusals.

The Claude 3 family consists of three models at different capability and cost levels: Haiku (fastest, most affordable), Sonnet (balanced), and Opus (most capable). All three are multimodal, accepting both text and image inputs.

Claude 3 Opus matched or exceeded GPT-4 on most benchmarks at the time of release. The report emphasizes not just capability but behavioral properties: the models are designed to be helpful, honest, and harmless while avoiding the problem of being overly cautious (refusing benign requests out of excessive safety).

The Claude 3 report advanced several important ideas. First, it showed that safety and capability are not in tension: the most capable model (Opus) was also the most nuanced in its safety behavior. Second, it introduced the concept of a "calibrated" model that neither refuses too much nor too little.

The tiered approach (Haiku/Sonnet/Opus) established a pattern now followed by most AI labs: offering models at different price/performance points so users can match the model to the task. Haiku for simple, high-volume tasks. Opus for complex reasoning.

Anthropic's focus on transparency in the technical report, including detailed benchmark comparisons and honest discussion of limitations, set a standard for responsible AI disclosure.

Authors: Anthropic (team paper).

Model family: Claude 3 Haiku, Claude 3 Sonnet, Claude 3 Opus.

Key advancement: Multimodal capabilities with strong vision understanding.

Link to paper: https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf

Full model card: https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf

Anthropic: Claude 3 announcement - https://www.anthropic.com/news/claude-3-family

Anthropic documentation - https://docs.anthropic.com/