Anthropic just admitted something that should make every operator pay attention

Jun 4 • AI news

This isn't a leak. It's not a rumor. Anthropic just published internal data on how much of Claude's development is now being done by… Claude.

The numbers are nuts. Read these slow 👇

→ 80%+ of all code merged into Anthropic's codebase is now written by Claude. (Before Claude Code launched in Feb 2025, that number was in the low single digits.)

→ Anthropic engineers now ship 8x as much code per quarter as they did from 2021–2025.

→ On a standard test — "take this small AI training script and make it faster" — a skilled human takes 4–8 hours to reach 4x speedup. Claude Opus 4 (May 2024) averaged ~3x. Mythos Preview (April 2026)? ~52x.

→ When shown a research session where a human made a wrong turn, Mythos Preview picked a better next step 64% of the time — up from 22% in 2024.

→ One April 2026 example: Claude shipped 800+ fixes that cut a class of API errors by 1,000x.

Engineering estimate of how long that would've taken a human: four years.

Anthropic's own framing: this is a possible path to recursive self-improvement — AI autonomously building a more capable successor. They say it's not here yet and not inevitable. But it's "happening faster than we thought."

My read on what this actually means:

The "AI builds AI" headline is the dramatic part. The part you should actually internalize is the rate of acceleration itself.

Look at the task-length curve they shared. Opus 3 (Mar 2024) could reliably handle 4-minute tasks.

Sonnet 3.7 (a year later) — 90 minutes. Opus 4.6 (a year after that) — 12 hours. Mythos Preview can hold a session for at least 16 hours and that's the top of what they can currently measure.

Task length is doubling roughly every 4 months. That used to be every 7. The curve is bending.

What this means for builders:

Whatever you couldn't automate 6 months ago — you can now. The "the AI can't hold a long enough chain of thought to do this end-to-end" objection is dying in real time. Workflows that needed three orchestrated agents last quarter can be one agent in 90 days. Plan accordingly.

What this means for operators:

The window where you can still get a meaningful edge from being early on AI is now. Not next year. Every quarter, the floor of "what AI can reliably do for a service business" jumps. The companies that build the muscle to deploy these capabilities the week they ship will compound on every release. The ones who wait for "stable, mature, proven" will be operating on Opus 4.6-era assumptions in a Mythos-era world.

What I'm genuinely thinking about:

The honest, slightly uncomfortable part. The same release that says "Claude writes 80% of our code" also says they don't yet know if Claude has the judgment to choose which problems matter most. That's still on humans. Which means right now, the operators and builders with real taste — who can pick the right problem to point an AI at — are the ones with leverage. That window is open. It won't be open forever.

When Anthropic themselves publishes the words "recursive self-improvement is plausible" — you don't

ignore it. You position for it.

If Claude could autonomously run ONE workflow in your business end-to-end for the next 6 months, what would you point it at? Drop it below 👇

5 comments

Anthropic just admitted something that should make every operator pay attention