DEDUPEDTop story: Meta's Muse Wonder
Meta Superintelligence Labs shipped Muse Wonder and Muse Wonder Full, its first frontier-class models, and the bigger one now sits second on Artificial Analysis' intelligence ranking, behind only Claude Opus 5.5. Both are closed, with no downloadable weights, a hard turn from the open Llama years. Wonder Full is locked inside $50 and $200 a month Muse Code plans, while the standard model is on Meta's API.
Here is our take: the benchmark is not the story. Wonder's reported API price roughly matches Opus 5.5, which beats it, so you are paying first-place money for second place. The real play is Active Memory, an agent that remembers you across chats. That is useful, and it also means Meta holds a detailed file on you.
What it means for you: no reason to switch if Claude or ChatGPT already works. Test it on one real task, and read the memory settings before you feed it anything personal.
Covered by 7 of 29
Also Big
Google's Gemini agent gets a job. Gemini Enterprise now acts like a coworker with its own email address, an audit trail and real-time spending caps, and it lets you swap in Claude as the model. Governance is the selling point, and we think that is what enterprises will actually buy. Covered by 4 of 29
OpenAI says it fired three safety researchers for leaks, not for speaking up. The company points to sensitive-information policy and disputes the researchers' public letter. Nobody outside can verify either side, so watch whether the third-party assessor contracts actually land. Covered by 2 of 29
Karpathy's "sharp funnel." He argues only about 5,000 people have true top-tier AI access, and that the gap between them and everyone else is widening. Treat it as a warning about who learns fastest, not a gloom story. Covered by 3 of 29
Microsoft's Maia 300 aims at Nvidia. Another hyperscaler chip, another attempt to cut the Nvidia bill. Real proof arrives when someone outside Microsoft runs production workloads on it. Covered by 3 of 29
Everyone Missed This
Epic told its Disney game team to write no code by hand. Bloomberg reports the project slipped from November to March, with developers told to go AI-first and make all art through AI workflows. The delay reportedly came before the order, so do not blame the robots yet. Still, a GDC survey found 52% of developers now call generative AI bad for the industry, up from 30% last year. Management is sprinting one way and its workforce the other. Covered by 1 of 29
Your Move This Week
Open the memory or personalization settings in whichever AI app you use most and read what it has saved about you. Delete anything you would not want in a breach. Ten minutes, and you will be ahead of the Muse Wonder memory debate.
Deduped. Every AI story. Once.
Know someone drowning in AI newsletters? Forward this to them.
