Morning Digest, September 23, 2026

9 newsletters, 9 overlapping stories

Top Stories

Meta’s Muse agent tops the App Store; Amazon blocks it, Shopify embraces it

(5 newsletters)

Meta’s Muse personal agent has overtaken ChatGPT at the top of the iOS App Store, reportedly logging 1.8M U.S. and Canadian downloads in 12 days versus ChatGPT’s 1.3M at launch, and Meta stock climbed more than 10% on the news. Amazon blocked Muse from making purchases on its site, citing privacy and security, though the move also protects a $68B ad business that depends on humans clicking. Shopify took the opposite bet, partnering with Meta to let Muse check out across every Shopify store via Shop Pay, setting up an early test of who controls agentic commerce.

Jev becomes Vercel’s fastest-adopted model as developers probe its limits

(5 newsletters)

TypeSafe AI’s Jev, a model that returns structured decisions and probabilities rather than generated text, was used by nearly 13% of paid Vercel AI Gateway teams within 24 hours. Ten thousand API probes suggest it is a causal decision model that shares state across questions, while a test on 100 real agent calls found it can struggle to beat a constant “benign” baseline, and one developer disputes TypeSafe’s “can’t hallucinate” claim. Open-source clones are already arriving: Kev offers small local Jev-style models at 0.8B to 9B parameters, and Tom Tunguz argues these models could cut if-statement costs by 99%.

OpenAI says its internal model has solved 100+ open math problems

(4 newsletters)

The same internal model behind OpenAI’s disputed Navier-Stokes claim has reportedly resolved more than 100 open problems across most branches of mathematics since it began training on Aug. 28. OpenAI formed an unpaid advisory group hosted at Princeton’s Institute for Advanced Study, including Timothy Gowers, Martin Hairer, and Melanie Matchett Wood, to advise on vetting and releasing results. The group has no say over pace, which matters because 27 Fields Medalists recently accused AI labs of rushing out results without proper credit or review.

Grok 4.7 posts big coding gains at the same price

(4 newsletters)

xAI’s Grok 4.7 is built on a larger base model with longer reinforcement training, adding self-verification, longer context, and better document and presentation output. Terminal-Bench nearly doubled to 38.0% versus 4.6, while pricing holds at $2 input and $6 output per million tokens. It is available through Grok Build, Cursor, and the API.

Anthropic ships Opus 5.5 as OpenAI halves prices with GPT-6 Sol and Luna

(3 newsletters)

Anthropic released Claude Opus 5.5, which takes the top spot on Artificial Analysis’s Intelligence Index at 58 (ahead of Fable 5.1 and GPT-6 Astra at 53) while costing 40% less than Fable, and it reportedly addresses the “Claudish” writing style by sticking closer to users’ style rules. Ninety minutes later, OpenAI launched GPT-6 Sol and Luna at half the price of the 5.6 versions ($2/$10 for Sol, $0.10/$0.50 for Luna), with Sol reportedly making about half as many factual mistakes. It is the first major launch day of the industry’s new “pacing” era, and it does not look much slower.

Xiaomi open-sources MiMo-V2.6 Pro and Flash

(3 newsletters)

Xiaomi released two MIT-licensed omnimodal models with 1M-token context, aimed at agent stacks: they can coordinate agents to build and visually test 3D scenes, generate Blender assets, and control a robotic arm from camera feeds. Flash delivers much of Pro’s agent performance at a fraction of the cost. Xiaomi also published the technical report, RL environments, and training code, and the Rundown puts the models roughly on par with Opus 5 and GPT-5.6 Sol.

Alibaba unveils the Zhenwu V900 AI chip

(3 newsletters)

At its Apsara Conference, Alibaba showed a homegrown accelerator it says triples the performance of its predecessor, plus plans for a 5 to 10 trillion parameter model and 20GW of data center capacity by 2032. Commercial availability is slated for early 2027, just as U.S. and Chinese officials negotiate over chip access.

a16z launches a college alternative for high school grads

(2 newsletters)

The Horowitz Andreessen Academy is a $42M, one-year, tuition-free program for about 50 students starting Fall 2027, admitted on what they have built rather than grades. Students get $50K in compute credits, co-ops at partner startups, and courses from figures like Sam Altman and Brian Armstrong, but no degree, since the program is unaccredited.

Google opens pre-orders for the Googlebook

(2 newsletters)

Google’s AI-first laptop line features a 2.8K OLED touchscreen, up to 14 hours of battery, Android phone integration, and Gemini features like Magic Pointer (summon Gemini anywhere on screen) and Rambler voice transcription. It is being called Android’s biggest move in years.

Also Worth Knowing

Quick Hits