AI Prompt
BlogAbout
← Back

Opus 5.5's Thinking Shift, Asahi Goes GPL, and More

4 min read

The loudest argument this weekend was whether Claude Opus 5.5 quietly changed. On Saturday, Shane Melaugh posted that he had Gemini 4 Argon read through 1,400 of his Claude Code sessions, and it flagged a sharp turn in how the model reasons starting November 3, with no change to the model name. The post spread fast, and a lot of people took it as proof of a nerf (Shane's post).

Theo dug into it a few hours later and came to a different conclusion. He says Anthropic appears to be rolling out a change to the reasoning summarizer account by account, and he saw no sign of degraded capability in his own work. His read is that Anthropic is cleaning up reasoning traces that have been messy for weeks (Theo's reply).

I think that's the more useful takeaway. What you see in the thinking output is a summary, so a change in its wording isn't the same as a change in the model. If you suspect a regression, compare finished tasks, cost, and turns on the same work before and after, not how the reasoning reads. Anthropic hasn't commented.

What Else Happened

Asahi Linux is moving all of its repositories from MIT to GPL-2.0-or-later. In its Linux 7.3 progress report on Saturday, the project said it had seen multiple attempts to hijack its work, including in-development code lifted from developers' personal forks and published without attribution, or "lazily rewritten with LLM assistance to obfuscate its origin." Going forward, all Asahi code will use strong copyleft licenses only. The report didn't say how existing contributions will be relicensed, and some developers have asked whether every past contributor agreed, since MIT code can't be relicensed without their permission. Either way, it's a clear example of AI-assisted code laundering changing how an open source project licenses its work. (Asahi Linux progress report, LWN, It's FOSS)

Garry Tan published open source evals comparing agent harnesses on the same models. His g-harness-evals project keeps the model fixed (Opus 5.5 or GPT-6.1 Sol) and swaps only the harness. Across 20 coding tasks, all seven setups passed 71 to 79% of the time, and the pass rates couldn't be told apart because 14 of the 20 tasks scored the same everywhere. Cost and speed did differ: on Opus, OpenClaw cost about $0.44 per solved task and took about 99 seconds per try, against $0.22 and about 40 seconds for Claude Code. On Sol, Codex was cheapest at $0.17. He says the tasks were too easy to separate the harnesses and a harder Terminal-Bench rerun is in progress, so I'd treat this as a cost comparison for now. (Garry's post, g-harness-evals on GitHub)

Brian Armstrong described how he now works with a team of agents. The Coinbase CEO runs one lead agent per area of focus, all the time, with a chief of staff agent overseeing them. Each lead manages sub-agents that write code, test it, chase reviews, and coordinate with people. Everything he needs to decide lands in one prioritized queue, a small local web app, and each product review comes with screenshots and a video demo. He says every change still goes through the same human review and security checks. The single queue is the part I'd copy, because switching between dozens of agent chats is where a lot of the overhead comes from. (Brian's post)

DHH says Omarchy will stay free. He wrote on Sunday that there's no monetization plan coming for his Arch-based Linux setup, and that the Omacom Foundation behind it is a nonprofit. His stated goal is to make Linux win on the desktop. (DHH's post)

Rohan Varma is leaving OpenAI. He announced on Saturday that he'd made the tough decision to leave, and thanked Alexander Embiricos for convincing him to join and Thibault Sottiaux for believing in him. He joined OpenAI's Codex team from Cursor. (Rohan's post)

Coming Up

OpenAI Developers posted a one-word teaser on Sunday: "Tomorrow." The short video points to a Monday announcement, and the account reposted Derrick Choi saying "the Chromatic" is coming to Supply Co. There are no details yet, so I'll wait for the actual launch. (OpenAI Developers, Derrick Choi)

What I'm Reading

  • g-harness-evals scoreboard – worth reading past the headline numbers for the notes on how each harness handles rules in the system prompt and skills it isn't told about
  • Asahi Linux 7.3 progress report – the licensing section is short, and the rest covers the project's hardware progress

That's it for today. 👋