Anthropic says the new model shows strong gains in coding. It can work more reliably across large software projects, review code, debug issues and even detect its own mistakes.