Anthropic Says Claude Writes Most Of Its Code Now. The Timing's Fishy

80% is a lot. That's the number Anthropic gave for a benchmark their internal coding agent is hitting. When Claude Code rolled out in February 2025, that figure was much lower. 5%. Not a lot. The company also added that engineers are merging eight times more code per quarter than they did in 2024, so currently at Anthropic, the human job is less writing and more directing Claude to write its own code.

The post wasn't just about stats though, it came with a warning. AI will soon be able to build its own successor. They're calling this ability "Recursive Self-Improvement". They're also asking for an industry slowdown so regulation can catch up. Very thoughtful. Why now?

We'll answer that, but first, some history.

Recursive Self-Improvement (RSI)

The phrase is stark. There's no hidden meaning, so you don't have to guess. It's used broadly in computer science and AI safety to describe this specific process of an AI system enhancing its own capabilities. Before RSI, there was "intelligence explosion", coined by British mathematician I.J. Good in 1965, describing how an "ultraintelligent machine" could trigger a chain reaction of self-improvement.

61 eventful years later, what Good described is happening faster than expected, and according to Anthropic, that's not… good.

The Acceleration

This is happening across the entire research loop, not just coding. In performance tests, each Claude model was given code that trains a small neural network and asked to make that training code run faster. The model has to find optimizations to do that.

These data points from Anthropic show how this is playing out:

Research Judgment: A major issue in AI research is knowing the best direction to follow. Mythos Preview was shown a session of a researcher making a wrong decision, and asked to make its own. It made a better choice of research direction 64% of the time. The number was 51% last November.

Task Horizon: This is fancy speak for the length of tasks AI can reliably complete, and it's expanding. The company says task length doubles about every four months, down from seven. In plain English, Claude Opus 3 could handle tasks equal to about four minutes of human work in March 2024. Same time this year, Claude Opus 4.6 was handling tasks that would take a human 12 hours. So the model can stay focused and reliable for longer periods.

The Open Research Demo: AI agents were given a problem in AI safety to solve independently. Two human researchers closed about 23% of the performance gap in one week, AI agents did it in 800 cumulative hours, using about $18,000 in compute to close 97% of the gap. So Claude is not just assisting human researchers but doing the research itself. It's this trend that Anthropic frames as a possible path to recursive self-improvement.

The Proposal

Anthropic believes a slowdown is needed so society can catch up. They will slow down, but only if everyone else does, and if there's a credible way to verify the global slowdown. They are currently researching possible methods. Very convenient. They get to look responsible while still racing ahead unless everyone else agrees to stop. Who's going to do that? Have you read about Sam Altman?

Why Now?

Here's why the timing is fishy.

The blog post came right after Anthropic reportedly submitted a confidential IPO registration, so all financial data, business strategies, etc. are hidden till much closer to the stock market debut. That means the competitors they want to "slow down" don't get to see any of the cards in their deck till launch, and the recursive self-improvement narrative reaches investors right on time.

Sneaky.

Anthropic has built its brand around safety, but we've heard that one before. In April, Anthropic announced that its Mythos model had become "too powerful" to release to the public, then a security startup (AISLE) found the same exploit that Mythos had found using eight small, cheap models. The gap wasn't as big as they said.

Weeks before the June 4 announcement, devs also reported that Anthropic's models were performing badly, accusing Anthropic of "nerfing" the models to cut costs. After denying, the company admitted a bug and recent changes to the model's "harness" were responsible.

The company has a history of presenting model capabilities that are disconnected from real-world user experience.

Is True RSI Even Feasible?

Critics don't think so.

What Anthropic is describing is a system that can improve its own software engineering, but "vibe-coding" has its limits. Their report stats are great, but for now, an AI that can write better code still needs a goal and a metric for "better". That goal is set by humans, and will be for the foreseeable future.

The company's very own Jack Clark has admitted that they "severely underestimated the scale of technological advancement, being ill-prepared to face the risks of runaway models". If this is coming from a founder, it implies they don't understand their own trajectory. There's a "trust me bro" vibe around the topic.

But…

"Recursive self-improvement" sounds like sci-fi (can't wait for Dune Messiah), but Anthropic's data is real. The trend is real. It's not just them either, competitors are racing too. In May, OpenAI posted a job for "preparations for recursive self-improvement", with a hefty paycheck. AlphaEvolve (Google) found a matrix multiplication algorithm better than anything since 1969. Whether RSI happens by 2028 or 2038, that ship has sailed.

Whatever happens, we'll let you know. If. Or when.