Claude 3: Why Anthropic's Model Family Became a Major AI Milestone

Anthropic's Claude 3 launch post https://www.anthropic.com/news/claude-3-family was significant for a few reasons beyond the benchmark performance. The three-model family structure, Opus, Sonnet and Haiku, was a clear statement about how Anthropic thought about the product market: different capability-cost tradeoffs for different use cases rather than a single model for everything.

The benchmark results placing Claude 3 Opus competitively with GPT-4 on a range of evaluations changed the market narrative. Before Claude 3, the AI model conversation was primarily a debate about OpenAI models with Anthropic positioned as the safety-focused alternative. After Claude 3, Anthropic was in the frontier capability conversation as well as the safety conversation.

The vision capabilities being included across the family extended Claude from a text model to a multimodal model at a time when image understanding was becoming a baseline expectation.

The model family design question the launch raises is worth revisiting now that every major lab has adopted some version of it. The right answer to which model to use for a given task is not always obvious and the cognitive overhead of model selection is one of the under-discussed costs of a multi-model world.

Do you prefer working with a model family that offers different capability-cost tradeoffs, or would you rather have one model that covers everything at a flat cost?