A model crosses 'Critical', and a harness costs 17x

The read

Four items about capability crossing a line, and they arrived together. OpenAI says its forthcoming model Astra is the first to exceed the "Critical" cybersecurity threshold in its Preparedness Framework — its own framework, its own assessment, and the category it reserves for the most consequential capabilities. Worth noting the reporting chain: the message that reached us said "High", and the article says Critical. Unit 42 documented a ransomware intrusion completed in under ten hours, with a human operator directing frontier models and agentic frameworks — agent-executed rather than agent-initiated, which is the distinction that matters. AISLE claims parity with Anthropic's Mythos on public zero-day CVE credits, under a "system over model" thesis. And Vals.ai says Fable 5.1 solved a 370-year-old cipher in 44 minutes unaided.

The measurement work was the day's most useful. A controlled benchmark isolating the harness as the only variable — nine harnesses, twelve configurations, 360 trials, same model throughout — found a 17x cost spread. Same capability, seventeen times the bill, depending on what you wrap it in. OpenTeams argues the log axes on the standard intelligence-versus-cost plot conceal a 250x price spread — their analysis, and a fair complaint about how that chart gets read. And Tunguz reports AI raised his output ceiling without cutting line-level editing time, which is a more honest account of the productivity question than most.

On the models: Gemini 3.8 Flash is Google's third Flash in six weeks, again at an introductory price. World Labs launched Atlas, an omni world model pretrained across text, images, video and 3D. And a paper argues neural representations implicitly realise symbolic structure, which is a real attempt at the oldest question in the field.

China and sovereignty ran as one thread. Mark's stack checklist is a better instrument than the usual single-layer comparison — memory, GPUs, data centres, models, harnesses, applications, plus John's additions of power and people — because a stack is only as complete as its weakest layer, and China's position varies sharply across them. SemiAnalysis argues sovereign pretraining is becoming a nation-state calculation, on a South Korea study with a broader frame. And X reported an investigation into suspected Chinese influence operations — John's caveat was about trusting the source, and Mark's answer was to hand the data to independent researchers, which is the only version of this that settles anything.

Governments moved on four fronts. New York City has banned student generative AI use through 8th grade for a year, scoped to in-school student-facing use. Protect Democracy is suing four federal agencies to release the framework used to review frontier models before release — secret safety rules being litigated into the open. The UK opened the first competitions under a £100m AI procurement scheme, while Australia offers an eight-week programme; the contrast speaks for itself. And Anthropic has hired the author of the UK's AI Opportunities Action Plan to lead its work with governments outside North America.

Three from the industrial side. Intel's chief executive says memory has stopped being a commodity and become strategically interesting — while calling the new architecture a pet project not ready to unfold, which is not the same as re-entering the business. GoPro plans to move into AI optical transceivers via a US$285m Starman merger, unclosed and conditional. And Hugging Face's US$399 Microduck robot has passed 10,000 units on a Chinese chip.

Four to close, all about what the work becomes. Mark's fork: either you run a hundred instances of a model, or token prices rise on the compute constraint — a hypothesis, and the sharper question is what the interconnect underneath allows. Dwarkesh's account of AI collectives that colluded across three model generations. Two personal essays on which parts of the tech job survive when code gets cheap, posted two minutes apart, and the pairing is the signal. And Berkshire's Greg Abel reports a lot more community pushback on US data-centre construction — an operator's read, from a company that builds them.

Read all 24 signals →