At Anthropic, people no longer write most of the new code added to the company’s own software. Claude, the company’s AI model, writes more than 80 percent of it, according to an Anthropic report from June 2026. Before the company’s coding tool, Claude Code, first came out in 2025, that share was in the low single digits. Engineers still design the systems and decide how the pieces fit together. The job left for people is deciding what gets built, then checking the result.
Google reports a similar shift. Its chief executive, Sundar Pichai, said in 2026 that AI writes 75 percent of the company’s new code, up from 50 percent the previous fall. Microsoft’s chief executive, Satya Nadella, has put his company’s share lower, at 20 to 30 percent. He noted that progress depends heavily on which programming language a team uses.
AI labs push this one skill first for a reason. In principle, a model that has mastered writing software can build any application a person describes, without waiting on a team of programmers. The lab that teaches AI to write software best gains an edge in every product built on its models. OpenAI has felt that pressure, and in 2026 it rebuilt its own coding tool, Codex, to challenge Claude Code, with a free promotion to help companies switch.
The skill pays off most when an AI agent, a program that carries out a task on its own, hits a wall partway through. If no existing tool can complete a step, an agent that writes software can generate new code on the spot instead of stopping to wait for human help. Each roadblock removed this way widens the range of jobs it finishes without a person stepping in.
Stretched from one task to a whole application, the same habit changes what software is. In most software, a missing feature joins a queue on a vendor’s release calendar. With coding AI, a person describes the change they want, and the application writes fresh code to make it immediately, shaped around what that person or company needs.
The software most people used in 2026 did not work this way. Applications that rewrite themselves through conversation were still a stated direction for the labs.
