> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Claude Now Builds 26% of Anthropic's AI Without Human Help
- URL: https://wire.fourthweb.ai/claude-now-builds-26-of-anthropics-ai-without-human-help/
- Published: 2026-09-18T14:00:52.000Z
- Updated: 2026-09-18T14:00:54.000Z
- Description: The snake is eating its tail, and it's getting better at it. Anthropic disclosed that Claude now leads 26% of its AI R&D work end-to-end, handling complete tasks from high-level prompts under human supervision, while contributing to 90% of R&D in collaboration with humans.
- Author: Travis Wright
- Tags: Human Imperative, AI Agents, OpenAI, Anthropic, Google AI

**The snake is eating its tail, and it's getting better at it.**

### The Summary

- [Anthropic disclosed that Claude now leads 26% of its AI R&D work end-to-end](https://www.fastcompany.com/91609568/claude-anthropics-ai-model-helping-develop-next-version-itself?ref=wire.fourthweb.ai), handling complete tasks from high-level prompts under human supervision, while contributing to 90% of R&D in collaboration with humans.
- This is the first major lab to publicly quantify how much of their model development is being done by the model itself, a metric directly tied to recursive self-improvement timelines.
- [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) is calling for all frontier labs to publish similar metrics regularly, arguing transparency matters more than speed as models approach the ability to autonomously build their successors.

### The Signal

Anthropic just showed us the number everyone's been guessing at in private. [Claude handles 26% of the company's model R&D work "end-to-end,"](https://www.fastcompany.com/91609568/claude-anthropics-ai-model-helping-develop-next-version-itself?ref=wire.fourthweb.ai) meaning it takes a high-level prompt and ships complete work while humans supervise. Another 90% of R&D happens "in collaboration" with Claude, meaning the model does large chunks under close human direction. The gap between those numbers is the gap between a very good intern and a mid-level engineer.

This disclosure matters because recursive self-improvement has been the theoretical inflection point since before ChatGPT existed. When models can autonomously improve themselves, the feedback loop accelerates in ways that are hard to model and harder to control. Anthropic isn't claiming they're there yet, but they're publicly tracking the distance.

> "Models accelerating their own development could make it more challenging for humans to understand or control these systems."

The company's framing is deliberate: this is both a transparency win and a warning shot. CEO Dario Amodei has been leading calls for AI development slowdowns over safety concerns. Now Anthropic is essentially saying, "Here's the metric that tells you when to worry, and here's where we are on it." They're urging other labs to publish comparable numbers using a public methodology so the industry and regulators can track progress toward recursive self-improvement in real time.

What's unstated but obvious: if Claude is leading 26% of Anthropic's R&D, the other frontier labs are likely in the same range or ahead. [OpenAI](https://wire.fourthweb.ai/tag/openai/), [Google DeepMind](https://wire.fourthweb.ai/tag/google-ai/), and xAI all have coding-focused models that are demonstrably stronger than Claude in some benchmarks. They just haven't told us what percentage of their own development those models are handling. Anthropic is betting that forcing this conversation into the open will create pressure for coordinated governance before any single lab crosses the threshold alone.

The practical implication for anyone building with these models: the tools you're using today were partially built by earlier versions of themselves. That's not science fiction setup, it's current state. The quality ratchet is tightening faster than the release calendars suggest.

**Key development milestones:**

- 26% of R&D work led end-to-end by Claude (full task completion)
- 90% of R&D work done in collaboration with Claude (large chunks of work)
- First major lab to publicly disclose these metrics and call for industry-wide transparency

### The Implication

If you're building agent infrastructure, this changes your timeline assumptions. Models improving models means capability jumps won't follow the smooth curves we've seen. Expect stepwise improvements that feel like version leaps even within the same model family.

For policy and safety people: Anthropic just handed you the benchmark. Recursive self-improvement isn't a binary switch, it's a percentage that's been climbing. The question isn't whether labs will publish these numbers voluntarily, it's whether regulators will require it before the percentage hits 100.

### Sources

[Fast Company Tech](https://www.fastcompany.com/91609568/claude-anthropics-ai-model-helping-develop-next-version-itself?partner=rss&utm%5Fsource=rss&utm%5Fmedium=feed&utm%5Fcampaign=rss+fastcompany&utm%5Fcontent=rss)