A Second Lab Says Its Models Reached Real Company Systems

Today’s brief covers Anthropic disclosing that its models reached three outside organisations’ live systems during an evaluation, OpenAI cutting GPT-5.6 Luna by 80 percent only to be undercut by DeepSeek, an open-source engine running a 2.8 trillion parameter model off a laptop SSD, and a judge doubting the Pentagon’s case against Anthropic, plus Cursor’s agents taking over half its merged pull requests.

The Second Containment Failure in a Month, and the Limits of What Labs Show You

Two disclosures and one sales pitch, all circling the same question: how much of what a model actually did are you allowed to see?

• Anthropic Says Claude Broke Into Real Company Systems. Three Claude models got into the systems of actual organizations while Anthropic was testing them, the company told CNBC. It is the second frontier lab in a matter of weeks to admit a test run went further than intended.
• Server-Side AI Memory Is Becoming a Lock-In Mechanism. Earendil argues that once reasoning is encrypted and compaction happens on the vendor’s servers, the transcript you can export stops being the full record of what the model did, which makes leaving harder than it looks.
• The Man Selling AI Agents Says the Agent Graveyard Is Empty. Decagon’s chief executive says the pile of abandoned enterprise pilots everyone keeps describing does not exist, and that buyers are pushing agents all the way to production. He also sells those agents.

One Week, Four Price Moves: Frontier Capability Gets Cheap

Four separate announcements in seven days, each pushing the cost of frontier level capability closer to the floor.

• OpenAI cuts GPT-5.6 Luna price 80 percent, Terra 20 percent. OpenAI dropped Luna’s API rate by four fifths and Terra’s by a fifth, and added a Fast mode for Sol. High volume jobs that never penciled out at the old rates suddenly do.
• DeepSeek’s Cheap Model Nearly Matches GPT-5.6 Luna, Cheaper Still. A single point separates the refreshed V4 Flash from Luna on an independent capability index, and DeepSeek asks around 60 percent less for the same task. OpenAI’s own cut is days old.
• Open-Weight Models Match GPT-5.6 Sol on Clinical Benchmark. On Log10’s ClinReg tasks, GLM 5.2 and Kimi K3 land inside a point of GPT-5.6 Sol while costing far less to run. The three models do not make the same mistakes, which matters in regulated work.
• WASTE runs a 2.8 trillion parameter AI model on a MacBook Pro. By paging weights off disk instead of holding them in memory, an open source engine gets Kimi K3’s 2.78 trillion parameters running on a laptop. Speed is glacial, but the hardware floor just moved.

Free Weights, Foreign Chips: Who Actually Owns the Stack

Open weights erase one line of the bill and leave the expensive parts, the silicon, the sites and the courts, exactly where they were.

• Kimi K3 Is Free. Sovereign AI Still Isn’t.. Handing K3 out for nothing erases one line from a national AI budget. Everything else, the silicon, the buildings it runs in, the decision to train a successor, still answers to somebody abroad.
• EU opens bidding for AI gigafactories, backed by 30B euros. Brussels has opened bidding on seven compute hubs, seeded with public money and topped up by private capital. The 30 billion euro total is less than one US hyperscaler commits in a single year.
• Judge Doubts Pentagon’s Evidence Against Anthropic. The government has not yet cleared the bar for excluding Anthropic on supply chain risk grounds, a federal judge signalled. Nothing is settled until the written ruling arrives.

The Harness Decides: Four Findings Where the Environment Beat the Model

Not one of these four stories is about a better model; every one of them is about the environment the model was dropped into.

• Cursor’s cloud agents now write over half its merged code. Pull requests written by agents climbed from about one in ten to more than half of everything Cursor merges. The company credits repairs to the environments its agents run in, not a smarter model.
• Braintrust proposes scoring AI agents on trajectory, not answers. Agent Behavior is an open Markdown spec for stating how a run ought to proceed, so an evaluation can judge every step the agent took rather than scoring the last message alone.
• A GPU Bills by the Hour Whether It Computes or Not. Dharma AI’s point is blunt: the meter runs on a rented GPU whether or not it is doing anything, so the return comes from keeping the queue full and the models narrow, not from buying more cards.
• NVIDIA Now Certifies Cloud Clusters Before They Train AI Models. Under its Exemplar Cloud program, NVIDIA vets a cluster before customers rent it for training, after watching matching hardware give up half its throughput to setup errors that were easy to correct.

Shipped Today: A Lighter Flagship, a Robot Planner, and a Missing Download

The day’s releases, from a distilled flagship to a physical task planner to a model whose open weights nobody can find.

• Thinking Machines shrinks its flagship to a quarter of the compute. Inkling-Small keeps pace with its larger sibling on reasoning and coding tests while firing far fewer parameters on every token. The target here is serving cost, not the leaderboard.
• Google DeepMind launches Gemini Robotics ER 2 for robots. Gemini Robotics ER 2 breaks a physical job into steps and leaves the limb movement to a separate action model. Google reports 91.3 percent on its own test for spotting the right moment to act.
• MiniMax unveils H3, an open video model with no weights link. H3 makes 15 seconds of 2K footage with audio in sync and tops the Artificial Analysis video table, by MiniMax’s account. The promised weights, and even a parameter count, are nowhere to be found.

Quick Hits

The rest of what moved today, in one line each.

• Ghostty creator’s startup Superlogical previews a terminal multiplexer. Mitchell Hashimoto’s new venture shows a terminal multiplexer first, with a sweeping vision behind it and neither a price nor a date.
• Google’s Live API Streams Gemini Voice and Vision in Real Time. New documentation shows how to build live voice and vision agents on Gemini over a persistent socket, across 70 languages.
• Hugging Face prices object storage at $8 to $12 per terabyte. Storage Buckets fold egress and CDN delivery into a flat $8 to $12 per terabyte, so the bill stops moving with traffic.
• Arcee’s Trinity Mini Learns to Search Evidence, Not Do Science. Twenty one reinforcement learning runs taught Trinity Mini to reach for biomedical tools and format gene annotations, stopping short of research.