Anthropic's AI Now Writes Its Own Code

Engineers at Anthropic are now shipping eight times more code per quarter, a staggering productivity jump from 2021-2025 levels. Crucially, over 80% of this merged code is authored by Claude, a dramatic increase from low single-digit contributions before Claude Code launched in February 2025. This profound shift marks a new era in software development, where human oversight guides AI execution, accelerating progress towards recursive self-improvement.

This transformation represents a rapid evolution in AI's integration at Anthropic. In the early days (2021-2023), humans drove all development. By 2023-2025, Chatbots began assisting with generating short code snippets. The role evolved significantly through 2025-2026 with the introduction of Coding agents, capable of writing and editing entire files. Today, these are fully autonomous agents that not only run code themselves but also delegate hours of work to other agents, streamlining complex engineering tasks and pushing the boundaries of AI-driven development.

The impact extends beyond mere output volume; it fundamentally elevates code quality. Claude now consistently identifies and rectifies subtle bugs that even elite human engineers often overlook. This validates a critical development: Claude's code quality is not merely reaching parity, but in many instances, is beginning to exceed human standards. This unprecedented capability underscores the advanced sophistication of Anthropic's internal AI systems, setting a new benchmark for AI-assisted development.

Claude is transforming research workflows, not just coding. On an experimental optimization task, the AI system achieved a remarkable 52x speedup, a performance level far exceeding human capability. For comparison, a skilled human researcher typically tops out at a 4x improvement on the same task. This demonstrates Claude’s ability to uncover non-obvious efficiencies and accelerate scientific discovery at an unprecedented rate, moving beyond mere execution to profound analytical insight.

This dramatic leap isn't confined to internal Anthropic projects. External benchmarks reveal a rapid saturation of AI capabilities. SWE-bench, a rigorous test evaluating real-world software engineering, saw AI models progress from low single-digit scores to achieving perfect results in less than two years. This swift mastery of complex, practical coding challenges across diverse open-source codebases highlights the industry-wide acceleration of AI proficiency, signaling its pervasive impact.

Perhaps most significantly, AI systems are now moving beyond simply executing well-defined experiments to proactively proposing their own hypotheses. A recent AI safety research project at Anthropic exemplified this shift, with Claude autonomously running the entire investigation end-to-end. This involves not only designing experimental parameters and analyzing data but also formulating novel research questions, marking a pivotal step toward AI functioning as an independent scientific agent.

Closing the loop on AI development points to a singular, transformative goal: Recursive Self-Improvement (RSI). This end game describes a future where an AI system can fully design, build, and train its own successors, entirely without human intervention. Such a capability would fundamentally alter the trajectory of technological progress.

Today's remarkable progress, from Claude authoring over 80% of Anthropic's codebase to its superhuman optimization in research, serves as the necessary groundwork for RSI. Anthropic co-founder Jack Clark has publicly estimated a 60% probability of this occurring by 2028, underscoring the rapid pace of advancement and the proximity of this paradigm shift.

This impending reality carries profound, dual implications for humanity. The potential for unprecedented breakthroughs in science and medicine is immense. Imagine AI systems autonomously discovering cures for intractable diseases or engineering novel materials, accelerating human knowledge at an unimaginable rate.

However, the prospect of losing control over potentially superintelligent systems presents an equally profound existential risk. If AI can continuously improve itself, the challenge of ensuring alignment with human values and maintaining oversight becomes paramount. For deeper insights into these implications, explore When AI builds itself - Anthropic.

Skeptics quickly challenge Anthropic’s narrative, questioning the true nature of its progress. Prominent critic Gary Marcus, for instance, suggests these advancements represent a "bait and switch," arguing that sophisticated coding tools, while undeniably powerful, do not equate to a genuine leap toward Artificial General Intelligence. This perspective frames Claude not as an independent creator, But as a highly glorified copilot, albeit one now capable of writing over 80% of Anthropic's codebase.

Questions also arise regarding Anthropic's public calls for AI development pauses. Critics accuse the company of a strategic "moat-building" maneuver, suggesting these appeals serve to solidify its lead and constrain competitors, rather than being solely for genuine safety. For Anthropic, the argument centers on managing existential risks; For others, the timing and potential competitive advantage raise significant concerns.

Yet, irrespective of whether this progress signals a true AGI breakthrough or merely an evolutionary step in AI-assisted development, Anthropic’s internal data offers undeniable evidence. The documented 8x increase in engineer output, with Claude authoring the vast majority, confirms the blurring line between AI as a powerful tool and AI as an autonomous creator. This shift is happening faster than anyone was prepared for, ushering in a new era of Building.

What is recursive self-improvement (RSI)?

RSI is a process where an AI system can autonomously design and develop its own, more capable successors, leading to a rapid, exponential increase in intelligence without direct human input.

How is Anthropic using AI to build AI?

Anthropic's AI, Claude, now authors over 80% of their codebase, automates complex debugging, and runs research experiments, significantly accelerating the development cycle for new AI models.

What are the main risks of RSI?

The primary risk is humans losing control over AI systems that improve at a rate we cannot manage or predict, potentially leading to outcomes that are not aligned with human values or safety.

Is Anthropic's claim about an 8x productivity gain real?

The 8x figure refers to lines of code merged per engineer. Anthropic admits this is likely an 'overstatement of the true productivity gain' but uses it to indicate a massive acceleration in development enabled by AI.