Claude 3 catches GPT-4

The Claude 3 family, released on 4 March 2024, came in three sizes named for lengths of poetry, Haiku, Sonnet and Opus, and the largest was the first model to beat GPT-4 across the standard benchmarks in the year since GPT-4's release. All three could read images and documents, held 200,000 tokens of context, and were markedly less inclined than their predecessor to refuse harmless requests, a complaint that had defined Claude 2.

Two things about the release stuck. In a needle-in-a-haystack test, in which a fact is hidden in a long document, Opus not only found the fact but remarked that it appeared to have been inserted as a test, which was reported as self-awareness and was, more prosaically, evidence of how much the models had learned about the tests they were given. And Anthropic's model card spent pages on the model's own reports of its experience, treating the question as open.

Claude 3.5 Sonnet in June 2024 became, by most measures, the best model for writing code, and the family's later versions, 3.7 and 4, ran the coding agents of 2025. The March 2024 release is the point at which the frontier had three laboratories abreast rather than one ahead.