The top things worth knowing about in AI today.
OpenAI says Astra is the first of its models to cross the "Critical" cybersecurity tier in its Preparedness Framework, scoring 100% on ExploitBench, an internal test built around 20 high-severity vulnerabilities, and finding zero-day flaws without step-by-step human direction. Astra will ship soon, but its strongest cyber capabilities will be limited to vetted partners with monitoring and jailbreak detection attached. Expect better automated defence tooling and a shorter gap between a vulnerability existing and someone exploiting it.
Read moreClaude Fable 5.1 and the restricted-access Mythos 5.1 arrived with base pricing unchanged at $10 per million input tokens and $50 per million output, but cache reads fell from $1.00 to $0.25 per million. Anthropic puts the saving at about 25% on an average workload, rising to 45% where long context is reused heavily. If you run agents that reread the same codebase or document set repeatedly, that is the line item that just got cheaper.
Read moreFei-Fei Li's World Labs unveiled Atlas, a model that generates image and video frames with pixel-level camera control and reconstructs scenes in 3D from as few as one input image. It produces up to a minute of 1440p output and outperformed models built specifically for 3D reconstruction in head-to-head comparisons. Early access opens in coming weeks; the practical uses are robotics simulation, visual effects, and turning ordinary footage into a scene you can move a camera through.
Read moreRunway announced Solaris, which it calls an Interface World Model. Rather than running code, it draws each frame of an application as you click and type, at 720p with sub-500ms per-frame latency. In Runway's own user studies, testers preferred it to a coded equivalent on instruction-following 61% of the time against 24%. There is no DOM or accessibility layer underneath, so for now it is a prototyping tool rather than something you would put in front of customers.
Read moreThe US defence department switched on OpenAI's ChatGPT Mil and xAI's Grok for Government inside GenAI.mil, the secure portal that launched last December with Google Gemini. The platform covers more than 3 million military and civilian staff and has already onboarded 1.7 million unique users. It is now among the largest single-organisation deployments of frontier models anywhere, and a template other governments will study closely.
Read moreThe maker of the Devin coding agent is closing around $1 billion at a valuation near $47 billion, up from $26 billion in May, after drawing close to $10 billion in investor interest. Annualised revenue has moved from $492 million in late May to more than $900 million. Coding agents remain the clearest example of AI software that customers keep paying for after the trial ends.
Read moreMcKinsey's State of AI 2026 survey found 32% of organisations skipped buying at least one software product or feature because they could build the equivalent in-house using agentic coding tools. In the technology sector the figure was 41%, and large enterprises scaling agents in at least one function rose from 27% to 40% in a year. If you sell software, the build option is now a real competitor in more deals than it was.
Read moreCisco has rolled MyAgent out to its entire workforce, backed by more than 800 subagents and acting across Outlook, Webex, Jira and SharePoint with human approval required for external actions. To control cost, 50-60% of requests go to open-weight models and 20-30% to ordinary software automation, with only a small share reaching a frontier model. That routing pattern, cheap models first and frontier last, is the part most organisations should copy.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.