The Claude 3 family
Anthropic released three models of one family, Haiku, Sonnet and Opus, and the largest beat GPT-4 on most published measures.
Why it matters
The frontier stopped belonging to one company, and choosing a model became a choice between price and quality inside a single family.
Three sizes with the same behaviour and different costs made routine what had been the exception: fitting the model to the task rather than taking the largest. All three accepted images. Anthropic published a model card with safety measurements alongside quality measurements. The Opus lead did not hold for long, but the fact that it existed changed how the industry talked about OpenAI's distance from the field.