26 February 2025
Claude 3.7: The AI That Shows Its Working
26 February 2025
Anthropic has released Claude 3.7 Sonnet, and the headline feature is something called extended thinking: for hard problems, the model can pause, reason step by step, and — unusually — let you watch that reasoning before it answers. Alongside it, Anthropic previewed Claude Code, a tool that lets developers hand whole coding tasks to the AI from their terminal.
Why "showing the working" matters
One of the biggest legitimate objections to using AI in business has been the black-box problem: an answer appears, but you've no idea how it got there. Visible reasoning doesn't solve that entirely, but it helps in two very practical ways:
- You can spot the wrong turn. If the model misunderstood your brief in step two, you'll see it — and fix your instructions rather than just distrusting the output.
- You learn to write better prompts. Watching how the AI interprets your request is the fastest prompting lesson there is.
The quiet story: AI for real work
The pattern across every lab this quarter is the same: less party trick, more workhorse. Models that think longer on hard tasks, tools that complete jobs rather than sentences. For marketing teams, that means AI is graduating from "write me a caption" to "analyse these twelve months of campaign results and tell me what to stop doing" — a far more valuable question, provided you check the answer.