Anthropic is drawing a striking line between early AI assistance and the next phase of model development. In a new account of how its own engineering workflow has changed, the company says its Claude models now author more than 80% of the code that gets merged into Anthropic's production codebase, a dramatic jump from the low single digits before Claude Code launched in early 2025.
The disclosure is part of a broader argument from Anthropic that AI systems are already accelerating the development of AI itself. The company says this trend is visible both in public benchmarks and in its internal data, with engineers now shipping far more code than they did in previous years. Anthropic says the typical engineer was merging about eight times as much code per day in the second quarter of 2026 as in 2024.
Anthropic describes a progression in how its teams use AI. In earlier phases, workers relied on chatbots for small code snippets. More recently, coding agents have taken on larger chunks of work, including editing files and handling tasks with less direct supervision. The latest systems can run code on their own and hand off parts of a job to other agents.
The company says this shift has changed not just how much code is produced, but how engineers spend their time. Rather than typing every line themselves, Anthropic staff now set goals, review output and direct Claude through more open-ended tasks. The company says that in some cases Claude has handled work that a human engineer would have struggled to complete quickly because it required holding a large amount of unfamiliar context at once.
Anthropic also says internal use of Claude has expanded beyond routine coding. In one example, the company says Claude produced more than 800 fixes in April 2026 that dramatically reduced a class of API errors. According to the engineer overseeing that effort, a human would likely have needed years to finish the same work.
Anthropic does not claim that AI has fully replaced human judgment. The company says Claude is still stronger at executing clearly defined tasks than at deciding which problems to tackle in the first place. That distinction, it says, marks the difference between today’s systems and a future in which an AI could autonomously design its own successor.
Still, Anthropic says Claude’s output has improved substantially. It points to falling rates of human intervention during coding sessions and says the model is now handling more open-ended tasks successfully. On the hardest tasks it measures, Claude's success rate reached 76% in May 2026, up 50 percentage points over six months, according to the company.
Anthropic also says its internal review process has changed. Code changes are now checked by an automated Claude reviewer for bugs, security issues and other defects before merging. In a retrospective analysis, the company says a full deployment of that reviewer would have caught about one-third of the bugs behind past incidents on claude.ai before they reached production.
On the research side, Anthropic says Claude has become much better at running experiments that have fixed goals, such as speeding up small model-training programs while preserving correctness. The company says the model’s performance on that task improved from about a 3x speedup in May 2025 to roughly 52x by April 2026. Anthropic also points to an internal open-ended research demonstration in which Claude-powered agents were given an AI safety problem and asked to work through it with limited human direction.
Anthropic says the pace of model improvement suggests AI systems are becoming more capable of doing the work that improves future AI systems. The company argues that this could bring major benefits in areas like science and medicine, but also raises the stakes around safety, oversight and control.
The central message is straightforward: AI is no longer just assisting Anthropic’s researchers and engineers. According to the company, Claude is now helping build the company’s own future at a scale that would have seemed far more distant only a few years ago.