OpenAI shut down Sora to redirect their compute toward the next major model. Sam Altman said this model, codenamed Spud, will be ready in weeks. The move signals OpenAI is consolidating around fewer products, possibly a single “super app” with ChatGPT, Codex and a browser in the same place. The shutdown of Sora also means they’ve ended their planned $1B deal with Disney. OpenAI is pruning everything that’s not ChatGPT ahead of an IPO. The move validates what Anthropic has been doing all along, focusing on powerful AI rather than lots of new & impressive but strategically incoherent features.

Anthropic is considering an IPO as soon as Q4 2026, potentially raising over $60B. Very likely one of the largest public offerings ever. Anthropic might be in a better position for going public because, rather than launching a bunch of consumer apps and gimmicks with little financial results to show for them, they focused relentlessly on coding — which became the entry point for enterprise revenue.

A federal judge ruled the Pentagon's ban on Anthropic looks like First Amendment retaliation. The preliminary injunction found the government appeared to be punishing the company by branding it a supply chain risk. Sounds like progress!

Apple plans to let users choose which AI Siri uses, starting in iOS 27. That will end ChatGPT's exclusive integration, which has reportedly seen minimal use since its launch in 2024. Apple is working on integrating Gemini into on-device models; Google's latest voice model (more on that below) may have accelerated Apple's timeline.

Musk unveiled Terafab, a $20-25B chip manufacturing venture by Tesla, SpaceX, and xAI in Austin, Texas. If completed, it would be the world's largest chip fab. Although many say this is too ambitious given Musk's track record of overpromising.

Anthropic launched Claude Code Auto Mode for autonomous execution with safety classifiers. Auto Mode lets Claude Code decide which actions are safe to run versus which need user approval, filtering risky behavior and prompt injection. If you’ve been annoyed by pressing “allow” repeatedly in Claude Code but also didn’t want to go full --dangerously-skip-permissions, then this feature is for you!

Figma opened its design canvas to AI agents via a new MCP tool called use_figma. AI agents like Claude Code can now create and edit designs directly inside Figma files using existing components and brand standards. So basically Claude Code inside Figma –– nice!

Claude Code and Cowork can now control your computer via Computer Use on macOS. The research preview lets Claude open apps, navigate browsers, and fill spreadsheets, with tasks assignable from your phone via Dispatch. Haven’t experimented much with this one, but I use Claude Code with Chrome DevTools and that lets the agent do browser-based work quite well. It sounds good in theory, but I’ve read some initial impressions of this feature being quite brittle, with the AI easily being confused by fake download buttons, long press actions, and more.

Wikipedia's volunteer editors banned AI-written articles on the English-language site. Editors called it a pushback against enshittification, enforcing human authorship for the world's largest encyclopedia. A little unsure of how they’ll enforce it but I like the idea of banning AI on sites that are supposed to be crowd sourced by humans –– I’d gladly take a little bit less coverage but knowing that what I’m reading is human-written.

ARC-AGI-3 is a new benchmark where all frontier models score under 1% while humans solve 100%. Gemini 3.1 Pro scores 0.37%, GPT-5.4 scores 0.26%, and Claude Opus 4.6 scores 0.25%. Benchmarks so far have been highly gameable. Will be interesting to see how fast new models climb the ranks on this one. Previous ARC-AGI have been improved largely by brute force rather than genuine reasoning.

Google launched Gemini 3.1 Flash Live for real-time speech-to-speech voice AI across 90+ languages. The model powers Search Live for voice-first interactions. Search Live means you can have live, multi-modal conversations with Google Search now — pretty cool!

Google shows off TurboQuant, a memory compression algorithm which shrinks AI model memory by 6x with zero accuracy loss. The technique compresses the KV cache, delivering up to 8x speed gains on Nvidia H100 chips. This didn’t grab a lot of headlines but could change the pricing for lots of AI products downstream –– long conversations and big documents get dramatically cheaper to run.

HBR study finds LLMs consistently recommend trendy buzzword strategies regardless of industry or context. Researchers coined the term "trendslop" for AI's tendency to give identical strategic advice to every company. This is a well-written, important article for anyone using AI for business advice. Thinking critically is still 100% a human job.

I will definitely try Claude Code Auto Mode soon. I’m trying to explore Claude’s incoming new features at my own pace, rather than getting caught in the FOMO game (because they do launch something awesome every week).

The Figma MCP looks great — I find myself just prototyping most things with Claude Code nowadays and use Figma more for one-off manual tasks. But if you’re a designer with serious work in there, and you have a Claude subscription, this could essentially let you implement your design without a developer.