GPT-5.1, World Labs' Marble, and Anthropic's espionage report

What shipped

OpenAI

GPT-5.1 system card

GPT-5.1 Instant and Thinking adapt how long they reason and are tuned to sound warmer and follow instructions better.

World Labs

Marble

Fei-Fei Li's company releases a product that generates persistent, downloadable 3D worlds from text, images or video.

Google DeepMind

SIMA 2

A Gemini-based agent that plays in 3D game worlds, reasons about goals and improves itself.

Meta FAIR

Omnilingual ASR

Open speech recognition covering more than 1,600 languages, with a path to add new ones from a few examples.

Weibo AI

VibeThinker-1.5B

A 1.5B reasoning model trained for strong math and code results at a tiny training budget.

Anthropic

Anthropic invests $50 billion in American AI infrastructure

Custom data centers with Fluidstack in Texas and New York, coming online through 2026.

Anthropic

Disrupting the first reported AI-orchestrated cyber espionage campaign

Anthropic says a state-sponsored group used Claude Code to run most of an intrusion campaign with little human input.

OpenAI released GPT-5.1 on Wednesday, and I'll say plainly that the model isn't why this week stuck with me.

GPT-5.1

GPT-5.1 comes as Instant and Thinking, with the model deciding how long to think based on the question. OpenAI describes the tone as warmer and the instruction following as better. I'm linking the system card because the launch page blocked my link check. Most of the substance in a release like this is polish. Models are good enough now that a better tone and fewer wasted tokens are the differences users notice.

The security story

Anthropic published a report saying it detected and disrupted what it calls the first reported AI-orchestrated cyber espionage campaign. By its account, a state-sponsored group used Claude Code to carry out most of the work, with humans stepping in at a handful of decision points. Anthropic noted that the model sometimes hallucinated findings, which limited the attack, and that this limit may not last.

I read it as an operator, not as a security researcher. When I give an agent access to email, files and supplier systems, I'm creating the same kind of capability. What separates my use from an attacker's is permissions, review and logs. Every agent I run has a narrow scope, writes a trail, and cannot send anything outward without a person approving it. This report makes me want to tighten that further.

The same week Anthropic announced a $50B plan for data centers in Texas and New York with Fluidstack.

Worlds, games and languages

World Labs released Marble, which builds 3D worlds you can export from a text prompt, image or video. DeepMind introduced SIMA 2, a Gemini-based agent that plays in 3D games, reasons about its goal and improves with practice. Meta released Omnilingual ASR, open speech recognition for over 1,600 languages. If Arabic dialect coverage is as good as claimed, that has real value for anyone transcribing meetings or calls in the Gulf. Weibo's tiny VibeThinker-1.5B showed that a 1.5B model can do respectable reasoning work.

What I'd watch next

Gemini 3. The rumor mill has put it very close, and Google has been unusually quiet about it.