In partnership with
Welcome, AI enthusiasts
Greg Brockman ended OpenAI's launch briefing on yesterday with a line nobody expected to hear this year. He told reporters we are now in the AGI era. The scores behind that claim are the strangest part of it, with perfect marks on tests the model before this one could barely touch. Let's dive in!
In today’s insights:
OpenAI drops GPT-6, AGI Arrives
Three AI Giants went Dark Together
Anthropic wants add-ons Rewritting Claude Code itself
Read time: 4 minutes
LATEST DEVELOPMENTS
AGI ERA
🚨 OpenAI drops GPT-6, AGI Arrives
Evolving AI: OpenAI calls Astra the most intelligent model in the world and the perfect scores back the talks.
Key Points:
Greg Brockman ended the launch briefing with "Welcome to the AGI era" and said calling Astra the first model of that era is reasonable.
Asked what he believed himself the OpenAI president answered that "for me personally, I do think we're there."
Astra came out of the largest training run OpenAI has ever done, with more than 100,000 GPUs at its Stargate site in Texas.
Details:
OpenAI shipped GPT-6 Astra on Thursday and put it well ahead of the model it replaces on every test the company published. Astra took 99.9% on ARC-AGI-3 where the older version scored 7.8%, and it swept ExploitBench with a perfect 100%. On the office work benchmark it doubled its predecessor to 41.4%, and it finished computer tasks in about 47% less time. The model also found two Chrome flaws nobody had reported and moved a bound on prime number gaps that mathematicians had left alone for over 80 years.
Why It Matters:
GPT-6 Astra takes an order once and finishes the whole job with nobody watching the screen. The model works for hours on its own. Autonomy is what people mean when they say AGI, and Brockman says the first version of it has now shipped. Handing over a whole job runs up a bigger bill than asking a question, and OpenAI priced Astra at two and a half times its last model.
Free email without sacrificing your privacy
Gmail is free, but you pay with your data. Proton Mail is different.
We don’t scan your messages. We don’t sell your behavior. We don’t follow you across the internet.
Proton Mail gives you full-featured, private email without surveillance or creepy profiling. It’s email that respects your time, your attention, and your boundaries.
Email doesn’t have to cost your privacy.
AI INFRA
🌐 Three AI Giants went Dark Together
Source: Quartz
Evolving AI: Claude, ChatGPT and Grok all stopped answering yesterday morning, some of them for hours.
Key Points:
Downdetector logged more than 35,000 US reports for ChatGPT at the peak, and each company gave its own account of what failed.
Google never confirmed a Gemini outage and Perplexity users reported none, so two of the biggest American services kept working through it.
Cursor went down with the rest, since the coding tool routes its work to the same models that had stopped answering.
Details:
SpaceXAI blamed an outage at its Memphis compute center for taking Grok offline from 6:30 AM Pacific time, and it apologized to unnamed compute partners as well as to Grok users. OpenAI described a routing error that left ChatGPT and Codex unavailable for roughly half an hour. Cloudflare drew most of the early blame and told reporters that its own services were running normally. Anthropic rents the Colossus 1 cluster in Memphis from SpaceX to serve Claude, at a reported 1.25 billion dollars a month. Anthropic identified the cause within 24 minutes, then called it an infrastructure issue and would not say whose machines failed.
Why It Matters:
Claude and ChatGPT now carry customer replies and production code inside ordinary companies every working day. PagerDuty puts an hour of downtime above 300,000 dollars for most organizations, so a morning like this one costs the economy hundreds of millions. More of them will start buying their own hardware and running these models in house, where a failure is at least theirs to fix.
Evolving AI: Claude Code users could soon rewrite the tool while their employer decides the limits.
Key Points:
GitHub has hosted the plan since September 3, when Anthropic asked developers whether the company should build it at all.
Function Hooks is what Anthropic calls the change, and its own issue title says add-ons would come out ten times more powerful.
One demo shows Claude writing and loading an add-on that strips secrets out of tool output before the model reads them.
Details:
Anthropic's proposal lets people write add-ons that pass a request down a chain, where any add-on in the line can change that request or refuse it before the next one sees it. An admin who strips a power out of the single object every add-on works through takes it away from every add-on registered after it. Nine short videos show add-ons redrawing the screen so that sensitive values stay blurred through a screen share until you hover.
Why It Matters:
Claude Code would answer to the person shaping it right up until the account holder above them says no. Anthropic's plan hands whoever pays for that account a switch that narrows what the tool will do for the person using it. Nobody working on a company account would see that switch move, or find out from the tool that a limit had been set.
👀 Click on the image you think is real
QUICK HITS
🔐 Palo Alto Networks acquired AI security startup Console for a reported $500M, folding agentic workflows into Cortex.
💎 Mira Murati's Thinking Machines is in talks to raise $1B at a $40B valuation, more than tripling its seed price with Accel leading.
🏗️ Crusoe raised $3B at a $30B valuation, tripling in 10 months, after signing a $13B cloud contract with Jane Street.
🌀 Wonderful raised $550M at a $5B valuation, more than doubling in six months, with Salesforce joining as an investor.
📈 Trending AI Tools
📝 Granola - AI notetaker that captures the real insights and turns every conversation into ready-to-share, action-driving notes*.
🔍 LoupeKit - See what any page is built with and how much of it is AI.
🎥 Screenify Studio - AI agent records polished product demos on your Mac.
☁️ Epho - Run Claude Code, Codex, or Opencode in the cloud on your own repo.
*partner link