Why Anthropic Just Endorsed a Plan to Pace AI

What Pacing the Frontier Actually Asks For

On July 28, a new open letter titled Pacing the Frontier went live, and it is worth reading carefully because the headline framing going around, "AI workers want to slow AI down," is not quite what it says. The letter's central ask is one sentence: it wants the US government to support an international effort to build the technical and governance tools needed to deliberately pace the frontier of automated AI development. The signatories are explicit that they are not asking for a pause right now, they are asking for the option to slow things down later, in a coordinated, verifiable way, before anyone actually needs to pull that lever.

That distinction matters more than it sounds like it should. A call for an immediate pause is a position companies can dismiss as impractical, and competitors can ignore while gaining ground. A call for the infrastructure to pace development on purpose, later, if the situation calls for it, is a much harder thing to wave off, because it does not ask anyone to slow down today. It asks for the tooling to exist before it is needed, which is the kind of ask that is easy to agree with in principle and much harder to actually build in practice.

What makes this different from the open letters this industry has produced before is who signed it and how fast the companies themselves followed. Reuters reported that both OpenAI and Anthropic released company-level statements backing the initiative within hours of it going live, not weeks later after a PR review cycle. That speed is itself part of the story. A statement from individual employees is easy to treat as a minority opinion inside a company. A same-day company endorsement is not.

Letters like this have a track record of fading fast, signed by a wave of employees, covered for a news cycle, then never mentioned again once the company itself stays quiet. This one is already behaving differently, because the companies did not wait to see how the letter landed before deciding whether to associate with it. That is the detail that pushed this from "interesting employee statement" to something worth an actual article rather than a footnote in a weekly roundup.

Who Signed, and How Fast the Count Grew

The letter opened with more than 1,100 verified signatures from employees at frontier AI companies, and that number kept climbing through the following day, with later counts running well past 1,200. It is a live petition, not a fixed one, so any specific number I write here will already be out of date by the time this piece publishes. What stayed consistent across every source I checked was the seniority of who signed first.

Anthropic CEO Dario Amodei signed personally, alongside co-founders Jared Kaplan, Jack Clark, Benjamin Mann, and Chris Olah. That is not a junior-engineer letter forwarded up the chain, that is the leadership of the company putting its name on the document from day one. OpenAI's list included chief scientist Jakub Pachocki and Ilya Sutskever. Google DeepMind's head of AI safety and alignment, Anca Dragan, signed as well, and Meta's chief scientist Shengjia Zhao rounded out the most senior names reported across the major outlets covering this.

Anthropic's own public statement connected the letter directly to its own research, saying it supports the petition and pointing to its research on recursive self-improvement, published the previous month, as the reason tools for deliberate pacing are needed now rather than later. That is the part of this story that actually matters for anyone who writes code with Claude every day, because it means the letter is not abstract policy positioning, it is tied to a specific, measurable finding about what Claude is already doing inside Anthropic's own codebase.

The Research Behind It: Claude Writing Its Own Code

The report the letter leans on is called When AI Builds Itself, published by Anthropic on June 4. The headline number is the one getting quoted everywhere: as of May 2026, Claude authored more than 80% of the code merged into Anthropic's own production codebase, up from low single digits in early 2025. That is a fast climb by any measure, and it is the number driving most of the coverage.

The report has more detail worth knowing than that one number. The typical Anthropic engineer was merging roughly eight times as much code per month compared to the 2021-2025 baseline before Claude took on this much of the workload. On the hardest, least-specified coding tasks Anthropic tracks internally, the kind with no clear spec and multiple reasonable approaches, Claude's success rate reached 76% in May 2026, a jump of 50 percentage points in six months. On a narrower benchmark for optimizing training code specifically, the most recent Mythos Preview model reportedly hit a 52-times speedup over its original baseline, compared to roughly three times for Claude Opus 4 on the same task. The report also mentioned Claude agents running around 800 hours of open-ended AI safety research experiments on their own, which is the detail that most directly connects to why "recursive self-improvement" is the phrase doing the heavy lifting in every headline about this letter.

None of these numbers are predictions. They are Anthropic's own internal measurements of what already happened between early 2025 and May 2026, which is part of why the letter reads as more urgent than the usual policy statement. It is not warning about a hypothetical future capability, it is describing a trend line inside one company's own repository and asking for the tools to manage where that line goes next. I wrote about a related Mythos-generation result, its cryptanalysis work on two cipher designs, when Claude Mythos results first surfaced, and this new letter reads like the policy side of the same underlying capability jump that piece covered from the research side.

What stands out reading the two together is that neither one is describing a single dramatic breakthrough. Both are describing steady, measured improvement across ordinary internal benchmarks, the kind of numbers a company tracks quarter over quarter rather than the kind it saves for a keynote. That is a more convincing signal than a flashy demo would be, because it is the sort of data a company would rather not have to publish at all if the trend were not real.

What This Actually Means for a Solo Developer Running Claude Code

I am not a policy person, and I am not going to pretend this letter changes anything about how I plan my week. But the 80% number is not abstract to me the way it might be to someone who does not spend most of a day inside Claude Code, because it describes, in a more extreme form, a shift I have already felt happening in my own workflow over the last year. The shift from typing code myself to directing an agent that writes most of it is the same one I wrote about when agentic coding turned developers into directors rather than typists, and Anthropic's own number is just that same shift measured at a scale I do not operate at.

The honest, non-alarmist takeaway for someone running a one-person studio is not that the model is somehow unsafe to keep using. It is a reminder about where my own review discipline actually sits, given how much of what I ship already starts as Claude-authored code that I read, test, and approve rather than write from a blank file. If Anthropic's own engineers are living with 80% of their merged code coming from Claude and are asking, publicly, for pacing tools around that trend, then the review step on my side, the part where I actually read a diff before I merge it rather than trusting a green test run, is not optional busywork. It is the exact seam this whole letter is pointing at, just at a much smaller scale than Anthropic's own codebase.

I also do not think this letter should read as bad news for anyone building with Claude day to day. A company publishing its own numbers this specifically, and then asking regulators for tools to manage the trend those numbers show, is a company being unusually transparent about a capability jump most of the industry would rather not measure this precisely in public at all.

Bottom Line

Pacing the Frontier is not a call to stop building with AI, and it is not a call for an immediate slowdown either. It is a request for the infrastructure to pace deliberate development later, backed at the company level by both Anthropic and OpenAI within hours, and signed personally by Anthropic's CEO alongside senior researchers across the field. The number underneath the headline, Claude authoring more than 80% of the code merged into Anthropic's own codebase as of May 2026, is the concrete fact worth remembering once the policy language fades from the news cycle.

For me, running a one-person studio almost entirely through Claude Code, the useful reaction is not anxiety, it is attention. The same trend Anthropic measured at its own scale is visible, in a smaller way, in how much of what I ship starts as a Claude-written diff rather than a blank file. That is not a reason to slow down. It is a reason to keep the review step honest, which was always the actual job anyway.