Claude Is Writing 80% of Anthropic's Code — And That Changes What CEOs Need to Plan For

If you're a CEO, a CIO, or anyone responsible for AI strategy at your company, you need to understand what that data actually means. It doesn't just signal that AI is productive. It signals that the pace of AI development itself is accelerating in ways that will reshape every competitive dynamic you're competing in.

For most of AI's history, humans made every decision: What should we build next? What's the problem we're trying to solve? Humans designed the systems. Humans ran the experiments. Humans interpreted the results and decided what to try next.

That's beginning to change. Anthropic is now delegating increasingly complex and open-ended work to Claude. The data is stark: in Q2 2026, Anthropic engineers merged 8x as much code per engineer per day as they did in 2024. The difference isn't that engineers became 8x faster—it's that Claude is writing the code, and engineers are directing and reviewing.

More importantly, the work Claude is handling is getting more complex. In March 2024, Claude Opus 3 could reliably complete software tasks that take humans about 4 minutes. One year later, Claude Sonnet could handle 90-minute tasks. A year after that, Claude Opus 4.6 managed 12-hour tasks. If that trend holds, Anthropic says, tasks that take a skilled human days could come into range by the end of 2026.

Here's where the stakes get real. Recursive self-improvement is the term for an AI system capable of designing and building its own successor—without waiting for humans to specify what to do next. We're not there yet, but Anthropic's data suggests the path toward it is accelerating.

Think about the skill ladder. Early in your career, you execute tasks someone else specifies: "Fix the export button." With experience, you're handed a goal and you design the approach: "Investigate why the network slows down under heavy load." At the most senior level, you decide which problems are worth solving at all: "What should we build next quarter?"

Claude is already performing well at the first two levels. It can write underspecified code without human guidance and match skilled humans at executing well-specified experiments. The gap? Choosing which goals to pursue—the most senior-level work. But that gap is closing. The evidence:

When an AI system can reliably perform the most senior-level work—deciding what's worth building and then building it—you've crossed a threshold. That system can then build its successor. That's recursive self-improvement.

Anthropic isn't calling for panic. They're calling for preparation. The company explicitly notes that recursive self-improvement is "not inevitable" — but it could come sooner than most institutions are prepared for.

What matters for your business: the speed of AI development itself will likely exceed the speed of human policy, regulation, and organizational adaptation. If AI can build AI faster than humans can build governance around AI, the implications are massive.

Anthropic has even put odds on it: the company's co-founder Jack Clark publicly estimated a 60% probability of recursive self-improvement by the end of 2028. That's not a research paper hypothesis. That's a business leader's bet on when this becomes operational reality.

If you're a CEO, here's what you should be thinking about right now:

Anthropic's point isn't that AI development is speeding up in the abstract. It's that AI systems are literally becoming their own developers. The trend has implications far beyond AI labs.

You can read Anthropic's full analysis on their institute website, including the internal metrics, benchmarks, and the exact timeline of capability growth.