Claude 4.6 codes for 15 hours without losing context: what's left for programmers?

Claude 4.6: 15 hours of coding without losing the thread – is this goodbye to code grinders?

Saying it bluntly is not alarmism: Claude 4.6's ability to write code for 15 continuous hours without losing context has made the debate about whether AI will replace programmers obsolete. The question now is how fast, and at what cost. Anthropic's model, according to those who have tested it, reads entire projects, infers their structure and generates code without needing to be told what to do over and over. But the devil, as always, is in the details.

The productivity leap that scares people

An experienced developer compares their work with AI: programs that once took weeks are now sketched out in hours, but only on the condition of supervising every decision. The model doesn't know by itself that a function must include protection against SQL injection, or that a certain design pattern doesn't scale. The metaphor of the architect who has to specify everything down to the foundations keeps coming up: without technical judgment, AI builds castles in the air. Others, however, point out that Claude is "precise and concise", able to swallow a folder with dozens of files and deliver a coherent diagnosis. Productivity multiplies, but foundational knowledge remains the limiting factor.

The myth of the average programmer

The impact on IT jobs is not speculation: it is estimated that out of every 100 computer science graduates, only 10 reach a level that AI cannot match. Mediocre profiles — those who copy and paste solutions without understanding them — are the first to fall. Add to this the pressure of a global workforce that, combined with AI tools, can deliver the same work for a fraction of the cost. The result is not so much the disappearance of the sector as a brutal polarization: either you are a senior architect who knows what they're asking for, or your job fades away.

The energy cost of perpetual code

AI running for 15 hours straight does not come free. A detailed analysis of energy consumption reveals that a continuous inference session on a GPU cluster uses between 500 and 1000 kWh, equivalent to the electricity consumption of a Spanish household over two or three months. And that's without counting data center cooling, which adds an extra 40% to 50%. The ecological footprint of this revolution is the elephant in the room nobody wants to see, and it will most likely make access to these tools more expensive in the medium term.

In short, Claude 4.6 shows that the boundary between the human and the automated is blurring at a pace not even the most optimistic predicted. But the competent programmer is not dead: they have become a translator of needs, a supervisor of exceptions and an expert in legacy systems. The rest — those who wrote code without understanding why — are living on borrowed time.