Claude

Anthropic announced Claude on the same day as GPT-4. It had been in private testing for months with companies including Notion and Quora, and it was, by the benchmarks of the time, roughly a GPT-3.5-class model with a longer memory and a manner that users described as careful. What distinguished it was how it had been aligned. Rather than rely only on human labellers ranking outputs, the model was given a written constitution, a list of principles drawn from sources including the UN Declaration of Human Rights, and trained to critique and revise its own responses against it. Humans supplied the principles; the model supplied the feedback.

Constitutional AI, described in a paper the previous December, was the first alternative to RLHF that a frontier laboratory shipped, and it made the values a model was trained on into a document that could be read and argued with. Later versions published the constitution and, in 2025, a longer statement of the model's intended character.

Claude 2 in July 2023 raised the context window to 100,000 tokens, Claude 3 in March 2024 caught GPT-4, and the models became, by 2025, the ones most used for writing code. The March 2023 release is on this timeline as the moment the frontier acquired a second laboratory with a stated method for making models safe.