Anthropic published a detailed blog post on Thursday arguing that artificial intelligence systems are accelerating their own development at a pace that could lead to “recursive self-improvement,” a scenario in which AI designs and builds its successor with little or no human involvement, sooner than most institutions are prepared for. This provocative claim positions the question of self-improving AI not as a distant theoretical concern but as an imminent challenge that demands urgent attention from researchers, policymakers, and society at large. The post reflects a growing sense within the AI industry that the technology’s trajectory may be steepening in ways that existing institutions are ill-equipped to manage.
The post, co-authored by Anthropic Institute researcher Marina Favaro and co-founder Jack Clark, draws on internal company data to illustrate the trend in concrete terms. According to the company, more than 80 percent of the code merged into Anthropic’s production codebase is now authored by Claude, its AI system, a dramatic increase from single digits before February 2025. The company further reported that its engineers are now shipping eight times as much code per quarter as they did during the period from 2021 to 2025. These figures offer a vivid illustration of how deeply AI has already become embedded in the process of building and refining the very systems that power it.
Beyond the raw measure of code output, Anthropic shared data demonstrating that Claude’s ability to handle complex, open-ended tasks has improved sharply over a relatively short period. On the most difficult engineering problems, specifically those lacking any clear specification, Claude’s success rate rose by 50 percentage points in just six months, reaching 76 percent in May 2026. Even more striking is the company’s benchmark measuring how effectively Claude can optimize AI training code, where performance jumped from a roughly 3x speedup under Claude Opus 4 in May 2025 to approximately 52x under Claude Mythos Preview in April 2026. This acceleration in the AI’s capacity to improve the efficiency of AI development itself lies at the heart of the recursive self-improvement concern the company seeks to highlight.
The impact of AI assistance extends to the productivity of Anthropic’s own researchers as well. The company disclosed that in a March 2026 internal poll of 130 researchers, the median respondent estimated that they were producing around four times as much output with AI assistance as they would without it. This self-reported multiplier underscores the extent to which AI tools have already transformed the daily work of those at the frontier of the field, lending further weight to the argument that the pace of development is being meaningfully amplified by the technology itself.
Notably, the post conveyed a degree of caution and even concern about the implications of these trends. “If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing,” the post states. Anthropic indicated that it would be willing to slow down or temporarily pause its development efforts if other frontier labs did so in a verifiable manner. At the same time, the company candidly acknowledged the competitive dynamics at play, noting that unilateral action would simply shift the lead to a competitor, a tension that lies at the core of the coordination challenges surrounding frontier AI development.
The timing of Anthropic’s post is significant, as it arrived just one day after OpenAI published its own public policy agenda on Wednesday. OpenAI’s agenda called on the U.S. federal government to “prioritize monitoring progress towards recursive self-improvement” and to strengthen the Center for AI Standards and Innovation as the primary institution responsible for frontier AI safety. Demonstrating its own seriousness about the issue, OpenAI has also posted job listings for a safety researcher focused specifically on preparing for recursive self-improvement, offering compensation between $295,000 and $445,000. The near-simultaneous attention from two of the industry’s leading laboratories suggests that the prospect of self-improving AI has moved firmly onto the agenda of those shaping the technology’s future.
In its post, Anthropic outlined three possible futures spanning a wide range of outcomes, from capability gains stalling at current levels to a scenario of full recursive self-improvement in which, as the company put it, “the pace of progress in AI development becomes determined entirely by the availability of compute.” To address the coordination challenges these possibilities present, the company said it plans to organize conversations with policymakers, researchers, and civil society in the coming months. Anthropic compared the verification problem to Cold War-era arms control treaties, drawing a thought-provoking historical parallel while emphasizing a crucial difference: the world does not have decades to build the necessary trust this time. This sense of compressed timelines lends urgency to the company’s call for proactive engagement and thoughtful governance.
As artificial intelligence continues its rapid advance, Anthropic’s post serves as both a data-driven snapshot of the current moment and a sobering invitation to grapple seriously with the trajectory the technology may be following. Whether the future unfolds toward the cautious scenarios or the more dramatic possibilities the company describes, the issues raised demand careful consideration from all those with a stake in the responsible development of AI. By sharing internal data and openly acknowledging the difficult trade-offs involved, Anthropic has contributed to a vital and increasingly pressing conversation about how humanity can navigate the profound implications of systems capable of improving themselves, ideally in ways that preserve meaningful human oversight and align the technology’s development with the broader interests of society.
Related reporting: Claude Completes First Formal Proof of Fermat’s Last Theorem.
