Welcome back! Three hikers had to be rescued from Mount Shasta after using Google Gemini to help plan their trip. The group reportedly started hiking at 3 AM, reached the summit at 7 PM, tried to descend in the dark, and then spent the night in a canyon before rescuers found them the next morning.
The sheriff’s office said Gemini advised them to bring far less food and water than they needed, which is a useful reminder that your chatbot should not be your wilderness guide.
Is This the First AI Model in a While That Actually Justifies the Hype?
GPT-6 Astra has officially humbled me.
A few days ago, I made a whole video complaining about how frustrating it's become to get a "new" AI model every other day that feels only marginally better than the last one. And then GPT-6 dropped, my feed filled up with people losing their minds over it, and after spending some time with it... they're right.
This one actually feels like a big leap.
The specs that matter:
Computer use is the biggest jump. OpenAI says GPT-6 Astra can complete computer tasks nearly twice as fast as GPT-5.6 Sol, and it's dramatically better at coding, navigating software, and handling long multi-step workflows.
1.05 million-token context window with up to 128K tokens of output. You can give it the equivalent of multiple books or a large codebase as input, and get back nearly an entire novel's worth of content in a single response.
First OpenAI model at the "critical cybersecurity capability" threshold—which, given everything we've been covering lately, means Astra is genuinely more capable at offensive cyber tasks than anything OpenAI has shipped. Standard access is restricted for that reason.
The catch is the price. $10 per million input tokens and $50 per million output tokens—2.5x the cost of GPT-5.6 Sol. And above 272K tokens of input, prices double again.
My POV: For most people, GPT-5.6 Sol is probably still the right default—it's cheaper, plenty capable for everyday work, and doesn't burn through tokens as fast. GPT-6 Astra earns its price specifically on the tasks where the difference actually matters: long agentic workflows, complex coding across a big codebase, running a computer end-to-end, or serious research where you're feeding in massive context.
But honestly, it's a nice reminder that every once in a while, a new model does come out that really justifies the hype. If you've been feeling the same fatigue I have about AI announcements lately, this one is worth actually spending some time with.
Hit reply and let me know: have you tried it yet?
— Matt
Apple’s Big Launch Is Almost Here
Via TechCrunch
Apple’s September event kicks off at 1 PM ET, and this year’s launch is shaping up to be one of its most anticipated in a while.
What to watch: The big headline is expected to be Apple’s first foldable iPhone, reportedly called the iPhone Ultra. After years of watching other smartphone makers push foldables forward, Apple may finally enter the category with a phone that opens into a wider, iPad-like display and supports side-by-side multitasking.
Reports suggest it could cost more than $2,000, which is painful, but apparently still easier to justify than strapping a $3,500 headset to your face.
Huawei and Xiaomi conveniently launched new foldables just days before Apple’s event. Huawei’s Mate XT2 starts around $2,980 and folds out twice into a tablet-sized screen, while Xiaomi’s new 18 Fold starts around $1,540 and uses an in-house AI processor that Xiaomi says beats Apple’s A19 Pro in responsiveness.
AI hardware angle: Apple is also expected to talk about its Gemini-powered Siri update, a refresh to its smart speaker HomePod, and possibly show off AirPods with built-in cameras. On that last one, the idea is that the cameras would act as “eyes” for Siri rather than a photo or video capture tool, helping Apple move toward more ambient AI without immediately triggering everyone’s smart-glasses trauma.
The bigger picture: Apple has not been leading the AI narrative the way OpenAI, Google, Meta, or Anthropic have. But Apple still has the installed base, brand trust, hardware discipline, and loyal following to make emerging tech feel normal once it gets the timing right. This event will be worth watching not because Apple is first, but because Apple is finally deciding which parts of the AI and foldable future are ready for the mainstream.
Seattle Times and Newsday Sue OpenAI and Microsoft
The Seattle Times and Newsday are suing OpenAI and Microsoft. Their case lands in the middle of a growing legal pileup over whether AI companies are simply learning from the web, or building products on top of unpaid journalism.
The claims: The publishers say OpenAI used their reporting without permission as training data and that its models can reproduce passages from their work in response to user queries. Microsoft is also named because Copilot is built on OpenAI’s technology. The suit follows similar complaints from The New York Times, Ziff Davis, Merriam-Webster, Encyclopedia Britannica, and nearly 400 local newspapers.
What they want: The publishers are seeking damages, but they are also asking for something much bigger: the destruction of copies of their works, training datasets, and AI models that incorporate them. That is an aggressive remedy, and one that gets at the heart of the fight over whether copyrighted news content can be used to build commercial AI systems.
No response yet: OpenAI and Microsoft did not immediately respond to The Verge’s request for comment.
The bigger picture: LLMs rely heavily on news sources to answer questions about the world, but publishers argue those answers can reduce the need for readers to visit their sites in the first place. That creates the central tension: where is the line between sourcing information and infringing on the businesses that produce it?
Turn Company Memory Into Agents
Via Libra
Libra is an AI agent platform that connects Slack, Gmail, and calendar data into a unified company memory. It routes incoming messages, leads, feedback, and bug reports to autonomous agents that own each workflow end to end.
How you can use it
Route messages and tasks to specialized agents
Spin up agents from plain-English instructions
Automate workflows across teams
Replace task tracking with outcome-driven agent work
Pricing: Free and paid
Route AI Requests Privately
Via TrustedRouter
TrustedRouter is an OpenAI-compatible API gateway that routes requests across 600+ AI models from 90+ providers without logging prompt or output content. It runs on attested infrastructure across GCP, AWS, and Azure, letting users verify that the live gateway matches a public source commit.
How you can use it
Route requests across hundreds of AI models
Use zero-data-retention routing
Add EU-focused routes and end-to-end encryption
Set up provider failover and bring-your-own-key configs
Migrate with only a base URL change
Pricing: Paid
Jobs, announcements, and big ideas
Meta launches Muse, a private personal AI agent available through WhatsApp and a new standalone app.
OpenAI ships ChatGPT Images 2.5, cutting generation time in half and adding a sketch feature.
An NYU mathematician accuses OpenAI of copying a rival's approach to the $1M Navier-Stokes problem.
Anthropic draws a class action lawsuit over allegedly deceptive Claude Max subscription pricing.
Google DeepMind unveils AlphaGenome Atlas, a map of all 9 billion human DNA variants.
Blackmagic rolls out DaVinci Resolve 21.1 with an AI assistant, photo support, and over 100 new tools.
HUGE updates from ChatGPT! Here are the three you need to know ⤵️
That’s a wrap! See you Friday for more.
