Three threads run through today’s 17 stories.
Locked Doors: Who Gets to See What
An alleged recording device, a government agreement and a new complaint about AI that refuses to help. Access is the question in all three.
- OpenAI engineer arrested over an alleged recorder in an executive room. Prosecutors allege the device sat behind a TV in a room he could not enter. He has not been convicted, and OpenAI says it found no sign that model weights, code or customer data were touched.
- Anthropic opens a Singapore office and signs a safety agreement. The agreement lists areas where the two sides plan to cooperate, from model testing to training civil servants. It carries no contract value and no deployment.
- A security startup says AI guardrails are blocking the defenders. Artemis Security says its Orion-1 beats Claude and GPT at defending networks, on a test Artemis wrote itself. The better evidence is a case where big models refused to look at attack records.
Marking Their Own Work: The Numbers Come From the Sellers
Most of today’s figures come straight from the companies that benefit from them. Where an outsider has checked, we say so.
- Gemini 4 Argon is live on Google Cloud, at half price through Nov. 30. Google has put the model in its business app and API. The only performance claims so far are Google’s own.
- Liquid AI’s 24B model runs 2B at a time and beats a bigger rival on tool use. Liquid AI ran every test itself. It says it leads a larger rival on five of eight tool-use tests, one of them a 65.01 to 65.00 tie, and trails it at coding and general knowledge.
- Ramp is selling the AI system it built after running up $1.5 million in AI bills. CEO Eric Glyman’s post on X is the only source for the 323 staff, 20,000 monthly feature requests and a 30% lower bill. The post does not say what the 30% is measured against.
- What Character.AI told its users about its bankruptcy, and what it left out. The company’s own post blames headwinds and says the app shuts if no buyer appears. It never says what a sale means for users’ chats.
- Indian chemicals startup Scimplify raises $25M as buyers look beyond China. The Bengaluru company says it runs at $200 million a year, a figure that comes from its co-founder. Hitachi’s venture arm led the round and Hitachi is also a customer. A $250 million valuation is a leak.
Opening Up: Agent Tools Now Let Other Software Take the Controls
Five releases hand more of the controls to other programs, other models and ordinary users, from NVIDIA’s chip scheduler to a robot lab’s beta.
- Anthropic lets Claude write the program that runs its other Claude agents. In a conference demo, the presenters said the new flows feature found more detail than a single agent did, at a much higher cost.
- NVIDIA opens up the software that shares its chips, with strings attached. All of Run is now released, but the licence is murkier than open source suggests and support is not included.
- OpenCode 2.0 lets other programs drive a coding assistant. One background service now lets scripts and other apps find and control the assistant, so it is no longer something you can only type at.
- CopilotKit wants your app’s users to train its AI agent. Its CLiF system learns from the agent’s own conversations today. The headline idea, watching how people do a job on screen and turning it into lessons, is still in early access.
- Physical Intelligence opens its robot models to outside developers. A beta lets companies upload demonstrations and tune the lab’s models on its computers. The announcement includes no performance numbers.
Today’s Quick Hits
- A voice model you direct like an actor, with stage directions in brackets. Fish Audio’s Drama 3 takes plain-text directions in brackets for how to say a line, but the company’s page gives no pricing, speed figures or independent tests of the voices.
- Prime Intellect buys Moondream, the small vision model maker. A tiny team known for a compact image-reading model now sits inside a company that rents out computing power. The price, the staffing and the model’s open future were all left unsaid.
- GPT 6.1 Sol now holds 500k of context by default, up from 272k. An OpenAI staff post says the larger working memory should mean long tasks get summarised mid-run less often, at no extra charge in Codex and ChatGPT Work.
- Zvi Mowshowitz says the AI consciousness debate runs on motive, not evidence. The blogger takes no side on whether AI is conscious. His claim is that confident voices on every side lack the evidence their certainty implies, and that incentives shape what they conclude.