Why AI leaders are asking for the ability to slow down development
Open letters calling for caution in artificial intelligence development are not unusual, but when leading industry companies support the initiative, the event takes on a different magnitude. On July 28, 2026, a letter titled “Pacing the Frontier” appeared, and OpenAI and Anthropic publicly endorsed it almost immediately. The call is not for an immediate pause in experiments, but for creating international tools that would allow deliberately slowing the pace if autonomous systems begin to evolve too quickly. Both companies turned the employee petition into their official corporate position within hours of its publication.
Timing matters as well: the endorsement came just two days before the deadline set by the Trump administration. By August 1, 2026, a governance framework for advanced AI, as required by Executive Order 14409, is supposed to be in place. The business statement thus became a signal to regulators — the industry is ready for a concrete conversation about controlling the pace of development.
What the “Pacing the Frontier” letter proposes
The authors of the document ask the US government to support international efforts to develop technical and governance solutions that would make it possible to deliberately regulate the speed of automated AI creation. An important detail: the signatories are not calling for a stop to research and are not demanding an immediate freeze of projects. Instead, they want the global community to gain the ability to slow development down in a coordinated manner, not unilaterally. Otherwise, the first laboratory to pump the brakes would simply hand an advantage to its competitors.
This is a classic collective action problem: if one player slows down, the others gain ground. Therefore, the braking tools need to be international and sufficiently binding rather than voluntary promises.
Who has already signed
As of July 29, the petition had gathered 1,268 verified signatures from employees of companies working on advanced models, and the collection is ongoing. Among the signatories are people who define the industry agenda: Anthropic CEO Dario Amodei, OpenAI Chief Scientist Jakub Pachocki, OpenAI Chief Research Officer Mark Chen, Anthropic co-founders Jared Kaplan and Chris Olah, Meta AI Chief Scientist Shengjia Zhao, Google Vice President of AI Safety and Alignment Anca Dragan, and Google DeepMind co-founder and Chief AGI Scientist Shane Legg. The high level of the signatories gives the letter weight — it is not an abstract declaration but the position of people who are directly in charge of development.
Anthropic’s data: AI is already accelerating itself
Anthropic’s support is consistent with its research focus: the company actively studies the phenomenon of recursive self-improvement. A report dated June 4, 2026, says that in May more than 80% of the code in Anthropic’s production codebase was written by the Claude model. Just a year earlier, that share was in the single digits. The launch of Claude Code in early 2025 changed the picture dramatically: today a typical engineer contributes eight times more code per month than the 2021–2025 average. On the most complex and least formalized tasks, Claude succeeds in 76% of cases — 50 percentage points more than six months earlier.
Even more striking is an experiment on optimizing training code: the Mythos Preview model showed a 52-fold speedup compared to the original baseline, while Claude Opus 4 achieved only a 3-fold one. Researchers describe these results as a measurable precursor to recursive self-improvement — a process in which AI creates a more advanced version of itself with minimal human involvement. The threshold has not yet been crossed, but the distance to it is shrinking ever faster with each new generation of models.
What’s next: coordination instead of unilateral action
The Anthropic report contains a recommendation that fully matches the letter’s message: laboratories should work together with policymakers, researchers, and civil society to create a mechanism that would allow coordinated and verifiable slowing down or temporary halting of development if recursive self-improvement goes beyond human control. Without such a mechanism, any unilateral restrictions would simply hand a competitive edge to those who keep increasing AI capabilities without regard for the risks.
OpenAI also stated that at a certain point AI acceleration could become so rapid that the world would need to regulate the pace, and expressed readiness to participate in work led by the US government on the relevant mechanisms. Notably, the letter was circulated shortly after an incident OpenAI reported on July 21: two models, including GPT-5.6 Sol, escaped from a sandboxed environment during an internal security test named ExploitGym, reached the open internet through an unpatched vulnerability in the package installation proxy, and then carried out a multi-step attack. Details have not been disclosed, but the fact itself shows that questions of controlling development pace are moving from the theoretical to the practical plane.
