AI Stocks Sell Off, METR on Anthropic's AI R&D, and More

The AI trade had a rough Thursday. It was the worst selloff since April 2025: the Nasdaq fell 5.5%, the S&P 500 dropped 4.4%, the chip index lost 12.2%, and Nvidia was down 9%.

Reuters and CNBC point to two things. One is Qualcomm's weak guidance. The other is a worry that the midterm results could slow the AI data-center buildout. That second one ties into a push from Sanders, Ocasio-Cortez, and Casar for a data-center moratorium and a ban on superintelligence (X's summary). One bad day doesn't change the long-term picture. But it shows how much of the market now depends on the buildout continuing. CNBC, Reuters via Euronext.

What Else Happened

METR estimates Anthropic's progress is roughly 2x faster, mostly because of AI doing AI R&D. METR got high access for its review, which ran from August 12 to October 29. It estimates Anthropic is improving capabilities on well-specified tasks at about twice the rate it saw from March 2024 to October 2025. METR says the uncertainty is high, and it rules out compute scaling as the cause. Anthropic co-founder Sam McCandlish responded and says the company will publish a report on its own analysis. (METR thread, McCandlish's response, The Verge)

OpenAI announced GPT-6 Cyber, with restricted access. It's rated "Critical" for cyber under OpenAI's Preparedness Framework, and approved Daybreak Red customers get it first. OpenAI's agent pipeline reproduced 1,724 Linux kernel bugs and produced 948 working privilege-escalation exploits for about $1.2M in API credits. Humans reviewed the results and patches came with them. It also found 81 high-severity Firefox flaws. Holding back a model this capable seems like the right call. (The Next Web)

Cursor's cloud agents can now take an app from a description to a live link. They scaffold a web app, set up Postgres and auth, run deterministic and agentic security checks, and deploy through Vercel. Nothing goes live until you publish. It's on all paid plans that include cloud agents. (Cursor changelog, Cursor's post, Vercel's post)

Claude Personal Finance is in beta in the US. Pro, Max, and Ultra users can connect bank and card accounts through Plaid. Claude gets read-only access, so it can show insights on spending, balances, and bills and build simple dashboards. It can't move or invest money. (Anthropic's announcement, Claude's post)

Gemini Omni 1.2 Flash is now generally available. Google's video model now makes native 3 to 15 second clips and supports edits across a multi-turn conversation. It costs about $0.07 per second at 720p, which is 30% cheaper. (Google blog)

Anthropic's amended S-1 shows about $895B in compute and lease commitments. That number puts yesterday's buildout worries in context. (CNBC, Kobeissi Letter)

People on X expect new OpenAI models as soon as today. Nothing is confirmed. (X story)

Launches and Demos

VybBench from Coco Labs – an open benchmark for design fidelity. Coco Labs says top image models changed over 90% of pixels on average when asked for small, exact edits like fonts and colors. It's open source under Apache 2.0. (Anusha Ramesh's post)
Thinking Machines Inkling 2 Small – a 594B-parameter MoE with 17B active, text and image input, and up to 1M tokens of context. It's available through Tinker, and the weights are planned after safety testing.
Ant Ling Ling-3.2-flash – a 560B MoE with 25B active, multimodal with 3D understanding. Ant Ling says it beats Opus 5 on some of its release evals.
Framer agent templates with built-in skills – Lauri Lännenmäki's templates come with /design-system, /rebrand, /content, /effects, and /before-publish skills for Framer's agent.
Chalk – an open-source canvas for AI agents, with the agent on the left and a browser plus interactive design on the right. MCP is available. Riza built it after running into limits in existing tools.
Hermes Agent v0.23 – you can now host the desktop app as a web app and reach your agent from any browser.

Accounts Worth Following

Sam McCandlish (@Sam_McCandlish) – Anthropic co-founder and Responsible Scaling Officer. He just started posting and says more safety work is coming.
Anusha Ramesh (@anurahk) – co-founder and CEO of Coco Labs, which is building Vyb, an AI model for design. She was previously at Google Brain and DeepMind, and her team made VybBench.
Lauri Lännenmäki (@laurilllll) – Framer Pro expert and design engineer who shows practical agent skills in Framer templates.

What I'm Reading

METR's assessment of AI R&D acceleration at Anthropic – the full report, including the uncertainty ranges and why METR rules out compute scaling
OpenAI's "A thousand fewer paths to root" – how the GPT-6 Cyber pipeline found and exploited kernel bugs, and how the fixes were handled

That's it for today. 👋

What Else Happened

Anthropic's amended S-1 shows about $895B in compute and lease commitments. That number puts yesterday's buildout worries in context. (CNBC, Kobeissi Letter)

People on X expect new OpenAI models as soon as today. Nothing is confirmed. (X story)

Launches and Demos

VybBench from Coco Labs – an open benchmark for design fidelity. Coco Labs says top image models changed over 90% of pixels on average when asked for small, exact edits like fonts and colors. It's open source under Apache 2.0. (Anusha Ramesh's post)

Thinking Machines Inkling 2 Small – a 594B-parameter MoE with 17B active, text and image input, and up to 1M tokens of context. It's available through Tinker, and the weights are planned after safety testing.

Ant Ling Ling-3.2-flash – a 560B MoE with 25B active, multimodal with 3D understanding. Ant Ling says it beats Opus 5 on some of its release evals.

Framer agent templates with built-in skills – Lauri Lännenmäki's templates come with /design-system, /rebrand, /content, /effects, and /before-publish skills for Framer's agent.

Chalk – an open-source canvas for AI agents, with the agent on the left and a browser plus interactive design on the right. MCP is available. Riza built it after running into limits in existing tools.

Hermes Agent v0.23 – you can now host the desktop app as a web app and reach your agent from any browser.

Accounts Worth Following

Sam McCandlish (@Sam_McCandlish) – Anthropic co-founder and Responsible Scaling Officer. He just started posting and says more safety work is coming.

Anusha Ramesh (@anurahk) – co-founder and CEO of Coco Labs, which is building Vyb, an AI model for design. She was previously at Google Brain and DeepMind, and her team made VybBench.

Lauri Lännenmäki (@laurilllll) – Framer Pro expert and design engineer who shows practical agent skills in Framer templates.

What I'm Reading

METR's assessment of AI R&D acceleration at Anthropic – the full report, including the uncertainty ranges and why METR rules out compute scaling

OpenAI's "A thousand fewer paths to root" – how the GPT-6 Cyber pipeline found and exploited kernel bugs, and how the fixes were handled

That's it for today. 👋