Claude 3.7 Sonnet and Claude Code
Anthropic released a model in which a fast answer and long deliberation are one mode with an adjustable budget, and with it an agent for programming in the terminal.
Why it matters
Reasoning stopped being a separate model: thinking longer became a parameter of the request, set by the user.
Until then the industry split into fast models and reasoning models, and the choice had to be made in advance. Here there is one pass, and the tokens spent thinking are given as a number, up to 128 thousand. The chain is shown. Claude Code shipped as a research preview: a program in the terminal that reads a repository, edits files, runs tests and makes commits. That moved the agent out of the browser into the environment where a programmer actually works, and within a year that way of working became the usual one.