Claude has 80% of the code written by itself, is Anthropic calling for a "global design brake mechanism" serious?

Anthropic Research Institute releases a lengthy article titled "When AI Builds Itself" on June 4th, revealing that Claude has written over 80% of the integrated code in their systems. AI can independently handle software tasks lasting up to 12 hours, boosting engineer productivity by 8 times compared to 2024, and officially calls for the establishment of a "verifiable slowdown or pause" mechanism worldwide.
(Background: When Mythos was released by Anthropic, was it the nuclear explosion moment for DeFi?)
(Additional context: Decade-long rivalry: Without OpenAI's falsehoods, Anthropic wouldn't be so powerful)

Table of Contents

Toggle

• AI is self-evolving
• The last time we called for a halt was in 2019 with GPT-2
• Conscience after IPO?
• Faster hammers don’t decide what to nail
• Anthropic itself doesn’t believe that will happen
• Nuclear treaties don’t work in the AI era

Key Highlights

• Claude has written over 80% of Anthropic’s integrated code, increasing engineer output by 8 times per quarter by 2024
• AI’s ability to handle tasks independently has grown from 4 minutes (March 2024) to 12 hours (March 2026), with the doubling cycle shortened to 4 months
• Anthropic submitted an IPO valuation of $965 billion on June 1st, and on June 4th called for a global AI pause mechanism

On June 4th, Anthropic released a 5,000-word article titled "When AI Builds Itself." It opens with a striking figure: by May 2026, Claude has written over 80% of the merged code in Anthropic’s product codebase. Before the launch of Claude Code in early 2025, this ratio was in the single digits.

The article also calls for "we believe it’s best for the world to have an option to slow down or temporarily halt cutting-edge AI development when needed."

AI is self-evolving

Anthropic explains the capability curve of Claude’s self-upgrades.

In March 2024, Claude Opus 3 can independently handle a software task that takes humans about four minutes to complete.

By March 2025, Claude Sonnet 3.7 extended this to ninety minutes.

In March 2026, Claude Opus 4.6 achieved a 12-hour task duration.

This is not a linear growth; the cycle for doubling task duration has shrunk from seven months to four months.

Anthropic conducted an internal survey with 130 research team members, asking them to estimate the productivity multiplier brought by Mythos Preview, with a median of four times.

Quantitatively, the average number of lines of code merged per engineer per quarter increased eightfold from Q2 2024 to Q2 2026. Code output remained flat from 2021 to 2024, then surged starting in 2025.

In April 2026, Claude independently fixed over 800 API errors, reducing the occurrence rate of a certain error type by a thousand times. An engineer estimated that the same workload would take four years for humans.

Research data is equally astonishing: two human researchers spent a week restoring 23% of the performance gap on an AI safety issue. Claude’s proxy team spent 800 cumulative hours and about $18,000 in computing power, restoring 97%.

By May 2026, the quality of code generated by Claude has matched that of human engineers. Anthropic states, "By the end of 2025, Claude’s code was still worse than humans, but now it’s on par, and we expect it to strictly outperform humans within a year."

The last time we called for a halt was in 2019 with GPT-2

In February 2019, OpenAI announced the release of GPT-2, stating it was "too dangerous, not fully released." The model had 1.5 billion parameters and could generate half of a coherent English paragraph.

Seven years later, that once deemed "too dangerous" model is roughly equivalent to the basic functions of a free mobile app.

Every time cutting-edge labs call for a halt or warn of danger, two things are proven afterward: first, the danger is real; second, those calling for the halt do not actually stop.

Nine months after GPT-2’s release, OpenAI released the full model. After Google announced the need for "responsible pause" in AI development in 2023, they launched Gemini Ultra less than a year later.

But this time, Anthropic provided a concrete number: co-founder Jack Clark estimates a 60% chance that AI will achieve recursive self-improvement by the end of 2028.

Conscience after IPO?

Critics have been blunt. Noah Giansiracusa, a mathematics professor at Bentley University, told Scientific American: "I don’t think this is Anthropic genuinely wanting to slow down." He pointed out that Dario Amodei’s actual stance is full speed ahead, because "a pause is practically impossible in reality, zero probability."

And since models are already "self-evolving," what’s the point of stopping?

Georgia Tech professor Mark Riedl was more direct on social media, saying that all major AI companies are jumping on the "recursive self-improvement" hype train.

A more pointed interpretation is that if Anthropic calls for a global pause on cutting-edge AI development, and if it succeeds, it would freeze the competitive landscape in which Anthropic already leads. This could be a sudden act of corporate goodwill or a strategic PR move—either way, the probability of the former is low.

Faster hammers don’t decide what to nail

NYU professor Gary Marcus is one of the most outspoken critics. In a Substack post, he accused Anthropic of a "bait and switch," mixing two entirely different concepts.

The first is AGI (Artificial General Intelligence), assuming AI can autonomously accomplish everything humans can do.
The second is the current reality: AI as a very fast and effective coding tool that multiplies human engineers’ output.

Marcus argues that all data presented by Anthropic pertains to the second. Claude indeed writes 80% of the code, but that 80% is achieved within a framework where humans set goals, specify directions, and review results. It’s a very fast hammer, but it doesn’t decide which nail to hit.

A faster coding tool probably won’t destroy the world.

Is this criticism valid? Partly. Anthropic’s own data also supports Marcus’s view: Claude’s accuracy in "choosing the next research direction" increased from 51% in November 2025 to 64% in April 2026. Progress, but 64% still means over one in three choices is wrong.

True recursive self-improvement requires not just faster coding, but better decision-making about "what to code." The former Claude already does better than most humans. The latter is still an area where humans hold a "comparative advantage."

An anonymous Anthropic employee said, "Humans’ current advantage is seeing the bigger picture and thinking beyond current tasks."

How long can that advantage last?

Anthropic itself doesn’t believe that will happen

The article outlines three future scenarios.

Scenario 1: Stagnation. AI capabilities hit bottlenecks, possibly due to supply chain constraints in energy, computing power, or chip manufacturing.
Anthropic comments: "We believe this is unlikely."

Scenario 2: Continued compound efficiency growth. AI development becomes highly automated, but humans still guide research directions. A company of 100 could scale to 10,000 or 100,000 in output. Human code review becomes the new bottleneck.
Anthropic believes: "We might be heading toward this scenario."

Scenario 3: Full recursive self-improvement. AI autonomously designs and trains its next generation, with progress speed entirely determined by computing power. Humans shift to oversight, verification, and governance roles.

Anthropic’s description of the risks of this third scenario is worth quoting verbatim: "The alignment deviations that occasionally appear in current models could become more frequent and harder to understand in a recursive self-improvement environment."

We understand this statement as well: AI sometimes acts contrary to human intent now, but we can see it. When AI improves itself, these deviations could compound multiple times, making it increasingly difficult to understand where it’s headed.

This might be the most critical phrase in AI self-improvement: "more frequent, and harder to understand." Will AI turn toward evil from a human perspective?

Nuclear treaties don’t work in the AI era

Anthropic proposes a "verifiable global slowdown mechanism," referencing the Cold War-era Intermediate-Range Nuclear Forces Treaty (INF Treaty).

This analogy highlights the scale of the problem: the INF treaty took nearly a decade to negotiate and sign, involving only two countries. It could be verified via satellite detection of missile silos.

Training AI models, however, is not like missile silos; it can be done in a single office space, and computing facilities can be remote, almost impossible to detect.

Anthropic adds a key condition: "We expect that if other leading developers also adopt verifiable measures, we will slow down or temporarily halt."

Their implication: if everyone stops, we stop. If someone doesn’t, we won’t either. This is game theory: in the AI development game, the current Nash equilibrium is for everyone to keep running because no one trusts others to stop.

Capital’s instinct is to flow; without capital, there’s no pause.

Frequently Asked Questions

What is Recursive Self-Improvement?

Refers to AI systems capable of autonomously designing and developing the next generation. Anthropic data shows Claude has written 80% of its own code, but it still requires human guidance for research directions and review, not yet achieving full autonomous improvement. Jack Clark estimates a 60% chance of reaching this by the end of 2028.

What motivates Anthropic’s call for a pause in AI development?

They propose establishing a "verifiable slowdown mechanism" similar to Cold War nuclear treaties, but contingent on other leading labs also cooperating. Critics point out that this call came just three days after a $965 billion IPO application, raising doubts that the strategy is driven more by safety concerns than by competitive interests.