Claude builds Claude, ChatGPT moves into Word, Grok takes notes.

Anthropic published three new measurements of how much AI is doing inside its own lab. The headline: as of August, Claude leads 26% of Anthropic's AI research and development work, meaning it can take most of a task from a high-level prompt to done while a person supervises. In February, that share was under 1%. More than 90% of the work now involves Claude at least as a close collaborator.

Read the method before you panic or celebrate. Anthropic sampled 20% of its research staff every week in July, broke their work into roughly 15,000 tasks across 542 categories, and rated each task on an automation scale developed by Epoch AI. Claude helped with the rating. Anthropic says so, calls the numbers self-measured, and says it will embed independent third-party evaluators to check its work. The Associated Press covered the release the next day.

Two smaller numbers from the same page stuck with us: about 30,000 AI agents were doing research and engineering work on Anthropic's internal platform, and about 6% of its AI R&D compute went to safety work in the week it measured.

On September 17, OpenAI's ChatGPT release notes announced ChatGPT for Word: a sidebar that drafts from notes, summarizes documents, revises selected text and fixes headings and formatting. It is available on every ChatGPT plan, including Free, and joins the Excel and PowerPoint versions of the same Microsoft add-in.

The less glamorous item on the same page matters more if your company built its own GPT. On September 11, OpenAI said it plans to retire custom GPTs across ChatGPT plans and offer a migration path to plugins, which pair reusable instructions with connected apps. Existing GPTs keep working until the retirement date that applies to your account.

On September 18, xAI, which now publishes as SpaceXAI, released Grok Voice Transcribe 2.0, a speech-to-text model it says is twice as accurate as version 1.0 at the same price: $0.10 per hour of recorded audio and $0.20 per hour for live streams. The company says it ranks first for accuracy among 32 streaming models on the Artificial Analysis leaderboard. It labels each speaker, accepts up to 100 custom terms per request and can drop every "um" and "uh" from the transcript.

Atlassian says Loom now uses it to transcribe videos. That tells you where this is going: every sales call, demo and webinar becomes searchable text for pocket change.

Put the week together and a pattern shows up. AI is moving into the places work already happens: your documents, your calls and, at Anthropic, the lab's own research. A good model is no longer an advantage, because your competitor has the same sidebar on the Free plan.

What still separates one business from another is what you feed the machine: first-party data, recorded customer language, proof you can show and a point of view worth quoting. The companies that pull ahead over the next year will have the best inputs, not the most subscriptions.

If you want help sorting which of these shifts matter for your business this quarter and which ones are just this week's noise, that is the conversation our founders enjoy most. Start with the free plan below. See you in the next edition.