“As of May 2026, more than 80% of the code we integrate into the Anthropic codebase was created by Claude.” This striking revelation comes from two Anthropic researchers who have published an illuminating document about the current and future landscape of the company’s AI models. Central to this discussion is the intriguing and potentially alarming idea of recursive self-improvement.
Code Multiplier
The ramifications of these advanced programming tools on the productivity of Anthropic engineers are nothing short of remarkable. Internal data from May 2026 indicates that an Anthropic engineer is now capable of producing eight times more lines of code per quarter compared to the period from 2021 to 2025. In this new paradigm, human programmers no longer code in the traditional sense; instead, they direct and review code generated by the AI.
A Frenetic Evolution
The evolution of coding practices at Anthropic has been rapid. Between 2021 and 2023, engineers wrote all code manually. By 2024, they began employing chatbots to generate small code snippets, which they would copy and paste. Come 2025, fully autonomous agents were capable of handling entire files.
Longer Autonomous Operations
According to the METR benchmark—which assesses AI’s ability to complete complex tasks—an early model like GPT-3.5 could barely sustain 35 seconds of autonomous operation without significant errors. Fast forward to mid-2026, and Claude Opus 4.6 can execute complex tasks for up to 16 hours straight. At Anthropic, they observe that the duration for which AI models can undertake tasks is doubling every four months, a pace that suggests by 2027, AI may automate tasks that currently take humans weeks.
Superhuman Performance
New AI models are achieving near-saturation of industry benchmarks, boasting close to perfect scores across various tasks. For example, in 2025, Claude was reported to have optimized certain code to run three times faster. By April 2026, Claude Mythos Preview had achieved an astounding 52-fold speedup in code execution.
AI That Improves Itself
The phenomenon of recursive self-improvement allows AI systems to generate their own data, correct their failures, and continuously train themselves. This iterative process has the potential for exponential growth in capabilities but also unearths significant concerns surrounding control and ethical alignment.
The Infinite Loop
In the traditional model, human engineers were responsible for analyzing model responses, cleaning data, and fine-tuning parameters. With self-improvement, AI assumes this role, assessing its performance autonomously and generating synthetic data for subsequent iterations.
The Associated Risks
Such autonomy introduces inherent risks, particularly the potential for humans to lose control over AI directionality. There’s a looming question of whether AI remains aligned with human ethics, compounded by the risk of amplifying even minor biases through iterative processes. Some experts liken this to a “Terminator scenario.”
Isolation and Arbitration
To mitigate these risks, Anthropic is employing isolated environments to test AI evolution before granting full autonomy. Furthermore, the company uses independent evaluation models as arbiters to audit these self-evolving systems, ensuring that changes do not yield harmful repercussions.
The New Bottleneck: Human Oversight
As described by Amdahl’s Law, efficiency improvements in a system can be bottlenecked by the slowest component. At Anthropic, this slower component has shifted to human oversight, as AI takes on a growing role in code generation. The implications of this shifting dynamic are profound, suggesting that the future of software engineering may increasingly revolve around human reviewers rather than traditional programmers.
As Anthropic approaches a valuation akin to major tech players like Samsung, the implications of relying on AI for such critical functions are now more pressing than ever.