On March 4, 2024, Anthropic released the Claude 3 model family, comprising three models: Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku. In its announcement, Anthropic said the family “sets new industry benchmarks across a wide range of cognitive tasks,” with Opus showing “near-human levels of comprehension and fluency on complex tasks.” Anthropic described Opus as a model that “outperforms its peers on most of the common evaluation benchmarks,” and called Haiku “the fastest and most cost-effective model on the market for its intelligence.”
The Claude 3 models added “sophisticated vision capabilities on par with other leading models,” meaning they could interpret images alongside text. Anthropic also reported a “twofold improvement in accuracy” over Claude 2.1 on factual questions, that the models were “significantly less likely to refuse to answer prompts,” and that Opus achieved over 99 percent recall in a long-context retrieval test. The models launched with a 200K-token context window.
Claude 3 mattered as a marker of arriving multi-frontier competition: the first time models from a lab other than OpenAI were widely measured at or above GPT-4 level, on the strength of Anthropic’s own published benchmark claims. The three-tier Opus/Sonnet/Haiku structure also set the naming and segmentation pattern that later Claude generations followed.