Short notes and things worth keeping

GPT-5.6 feels like it came out yesterday, even though it’s been about two months. Anthropic released Opus 5.5 yesterday, and now today’s release of OpenAI’s GPT-6 Sol and Luna feels like a strong answer. Sol and Luna cost less on the API than their 5.6 predecessors and feel faster in my early use. Opus 5.5 and both new GPT-6 models have been excellent to work with. Watching both companies improve their models at this pace is very exciting.

My initial take on Opus 5.5 is that it is (thus far) a super fast, super affordable, well-needed upgrade over Opus 5! I highly recommend giving it a spin.

OpenAI’s GPT-6 Astra is rolling out to the public, and I’m excited about this one. After using it across three projects, it’s the best experience I’ve had building software with an AI model. It’s still early, but if you have the opportunity to try it, I highly recommend doing so. I wrote up my first impressions in the Journal.

Voice input to AI feels human until I reread the transcript. The emphasis I put into a sentence is gone… the model only gets flat text. But maybe that loss points past AI. How much human conflict comes not from what we said, but from what we assumed the other person would infer?

One of the quiet gifts of working with AI is that I can step in and out of a project without imposing on anyone. If I need to pause, the session just waits. No one is blocked, annoyed, or spending their day in limbo. That changes the emotional shape of creative work: I can follow the project at the pace my attention allows, then return without apology and keep going.

Anthropic being ordered to suspend Fable 5 and Mythos 5 access over a narrow jailbreak claim is exactly why AI governance needs clear process, evidence, and proportional remedies. My worry is the precedent: if this standard holds, every frontier model and agentic coding tool is one disputed demo away from abrupt recall. Emergency authority may be necessary someday, but vague directives can become policy by shock instead of policy by proof.

Anthropic doubling Claude Cowork’s 5-hour usage window for a month is useful, but the unchanged weekly cap matters. It means more room inside a focused work block, not more total weekly runway.