What happened in AI this week (Sept 1-9th)

Hey friends 👋

Everyone launched a model in the same week. Then half of them went down at the same time. Then OpenAI said it solved a century-old math problem. Here's the version you can actually use.

🚀 They all shipped in the same week. On purpose.

This was the headline. In about 72 hours:

• Sept 1 — Claude Fable 5.1 (and Mythos 5.1, locked behind a trusted-access program). Same price as Fable 5, but cache reads dropped 75%. Anthropic says typical work is ~25% cheaper, agentic work up to ~45%.
• Sept 3 — GPT-6 Astra. OpenAI's new flagship. Rolling out to Plus / Pro / Business / Enterprise. API is $10 / $50 per million tokens (input / output). First model OpenAI has rated Critical on their own cybersecurity scale, so some features are gated.
• Sept 8 — ChatGPT Images 2.5. Faster (up to 50% less wait than 2.0), better at keeping a face/product consistent across edits. Free tier included. API split: Flare (fast, default) and Sunburst (slower, campaign-quality).

My take: stop collecting models. Pick one for thinking (Fable or Astra) and one for pictures (Images 2.5). The rest is noise. I already posted the fashion-sketch → image thing on IG — Images 2.5 is the one that finally doesn't wreck the drawing.

💥 Thursday: ChatGPT, Claude, Grok, and Gemini all glitched

Sept 3. Morning. Four labs, overlapping errors, no AWS/Azure/Cloudflare outage to blame.

Claude Fable / Mythos / Opus threw errors around 9:23am ET. ChatGPT + Codex degraded by 10:43. Grok showed a "we're working on it" screen. Gemini spiked on Downdetector without a clean official post. Most of it was back by early afternoon.

OpenAI later pointed at a routing error. Everyone else was quieter.

My take: if your whole workflow lives in one chat window, Thursday was the reminder. Keep a second model logged in. Local (Ollama) is the boring backup that still works when the internet's models don't.

🧮 OpenAI says it solved a Millennium Prize problem

Sept 9. OpenAI published a Lean proof of Navier–Stokes — one of the seven unsolved Millennium Problems. They ran something like 10,000 agents for ~88 hours on Astra. They are not claiming the $1M Clay prize. Mathematicians have not signed off. There's already a credit fight in the comments.

My take: cool for the field. Useless for your Tuesday. Don't rebuild your stack around "AI did math." Wait for humans with chalk to argue about it.

⚠️ An Anthropic researcher walked out

Same day. Jacob Coxon quit and said the people building this stuff earnestly believe it could kill us all by the end of the decade — specifically the race toward self-improving agents.

My take: you don't need a doomsday takeaway. The practical version: agents that can keep going after you close the laptop are the new product category (Grok Bot, Claude's computer-use, Astra). Use them. Don't give them prod credentials on day one.

🛠️ Two things worth opening this week

• ChatGPT Images 2.5 — if you draw, design, or sell anything visual. Sketch → image is finally usable. Try: upload a fashion sketch, ask it to render fabric + lighting, then iterate without regenerating the pose.
• Google Pics (pics.new) — Google's Canva-shaped image editor, inside Docs/Slides/Drive. Object select, edit text in the image, upscale to 2K/4K. AI Pro / Ultra and most Workspace plans. If your team already lives in Google, this is the one that doesn't require a new tab.

Honorable mention from my camera roll: Dyson's AI toothbrush. A camera in a toothbrush. We are so back.

🎯 The one thing to remember

The labs synchronized the drop so nobody had a quiet week. That's a marketing calendar, not a reason to switch tools every 48 hours.

Learn one new surface this week (Images 2.5 or Pics). Ignore the Millennium Prize until a mathematician you trust posts.

See you next week 💛 Angel · [@creativelyange on TikTok & Instagram]

Enjoyed this? Forward it to someone who asked "wait, which model am I supposed to use now?"