January is already obsolete. My honest breakdown of Opus 4.6 + what it means for developers, leaders, and everyone in between.

Sixteen AI agents coded for two weeks straight — humans set the spec and validated the results, but didn’t write the code — and delivered a functional C compiler. Roughly 100,000 lines of Rust. According to Anthropic’s engineering blog, it compiles substantial real-world systems software — including the Linux kernel, PostgreSQL, FFmpeg, SQLite, QEMU, and Redis — and passes the vast majority of the GCC torture test suite. Cost: $20,000.

A year ago, autonomous AI coding topped out at about thirty minutes before the model lost the thread. Last summer, Rakuten got seven hours out of Claude and the engineering team passed the results around like a rumor. Seven hours was a breakthrough.

Thirty minutes to two weeks in twelve months. That’s not a trend line. That’s a phase change.

And it’s not just code. Rakuten put Opus 4.6 on their engineering issue tracker and it autonomously closed 13 issues and routed 12 more to the right team members in a single day — across a 50-person organization spanning six repositories. Anthropic pointed the model at open-source codebases with basic tools and it found over 500 previously unknown high-severity vulnerabilities in production software that human researchers and automated scanners had already reviewed. Two reporters with no engineering background sat down with Claude Cowork and built a project management dashboard in under an hour.

None of this was possible in January. Opus 4.6 shipped on February 5th. It has been less than a week.

Here’s what’s inside:

• The three-month gap. What changed between Opus 4.5 and 4.6 — a 5x context window, 4x retrieval improvement, nearly doubled reasoning scores, and agent teams — all in a single quarter.

• What working memory actually means. Why the MRCR v2 benchmark matters more than the context window number, and what it looks like when a model can hold 50,000 lines of code and actually know what’s on every line.

• The Rakuten proof. Production deployment data from a company managing 50 engineers across six repos with AI — not a pilot, not a demo.

• Team swarms. How sixteen agents coordinated to build a compiler using the same management structures that human teams use — and what it means that AI discovered hierarchy independently.

• Revenue per employee. The reported numbers from Cursor, Midjourney, and Lovable that suggest the relationship between headcount and output just broke.

• The honest pushback. What the skeptics are saying, why some of it is fair, and why this release is different from the ones that underdelivered.

• A personalized briefing prompt. Paste it into Claude and get a walkthrough of every Opus 4.6 change mapped specifically to your work, your tools, and how you actually use the model.

Let me walk you through what happened — and what it means for how you work, whether you write code or not.