Four threads run through today’s AI news, and the first is a government drawing a line. Washington has ordered federal civilian agencies to clear DeepSeek out of their systems, along with any vendor software built on it. China’s foreign ministry has objected.
The second is companies measuring themselves. MiniMax, Inception Labs, OpenAI and Meta each published their own scores, mostly with no named rival and no independent check, while a Contrary Research report asks whether the industry’s tests still measure anything.
The third is money in unusual shapes: Otter’s $320 million that buys the investor no shares, a $1.45 billion valuation for Eka’s robot hand, and Anthropic reportedly nearing a second $45 billion computing deal.
The fourth is a pair of quiet admissions. Meta’s agents got more reliable once the one in charge was given less to do, and Google’s Notebook now shows you the plan before it builds anything.
Who Sets the Rules: Washington, Bangkok and the Web Host
A government names a model it no longer wants inside its buildings, an Anthropic co-founder tells economists the singularity has begun, and a startup bets that the web’s next job is choosing which robots get in.
- US tells civilian agencies to drop DeepSeek, and software built on it. Homeland Security has ordered federal civilian agencies to stop buying DeepSeek and clear it from their systems. The order also reaches third-party products built on DeepSeek technology, so a vendor’s chatbot can trip it with no DeepSeek name on the invoice. It is not a nationwide ban, agencies can ask for exceptions, and China’s foreign ministry has objected.
- Anthropic co-founder tells the IMF the singularity has begun, and owns his role. Jack Clark told an audience in Bangkok that the singularity is here, pointing to AI help on open math problems and drug design, and called himself part of the cause. He said Anthropic’s own modeling puts US GDP at 44 trillion by 2030, with no currency given. The label is his argument, not a finding.
- A $2.2 billion bet on telling good robots from bad ones. Estonia’s Blackwall raised $100 million at a $2.2 billion valuation by letting web hosts resell its bot protection under their own names. Its pitch is that the job has shifted from keeping robots out to choosing which AI agents get in. TechCrunch’s report gives no accuracy figures for how it tells them apart.
Grading Their Own Homework: The Scores Nobody Checked
Several companies published scores for their own products, and one research group wants a test that holds up on a real job. A separate report asks whether the old tests ever steered anything well.
- Chinese AI lab reveals the cheap model it quietly ran in secret. MiniMax says an anonymous model that handled more than 200 trillion tokens of traffic was an early build of its new M3.1 Flash, now on sale at $0.15 per million input tokens. It claims agent costs 70% lower than its predecessor, which it does not name. Every score is its own, and the promised open weights have no date.
- A faster kind of AI text model scales up, its maker says. Inception Labs says Mercury 3, which drafts a whole answer at once and refines it, runs at 1,068 tokens per second, 4x to 8x faster than GPT-6 Luna and Claude Haiku 5.5. Those rivals are the fast tiers of OpenAI and Anthropic, not flagships, and every benchmark in the post is Inception’s own.
- OpenAI says its new models can finally handle the code behind cash machines. OpenAI says its latest models can migrate COBOL, the decades-old language still running banks: one scores 86.6% on a byte-for-byte test where its predecessor managed 5.4%. The exams are OpenAI’s own and the comparison is to its older models, though it is publishing the setup for one of them, MultiLangBench, so outsiders can rerun it.
- The tests AI labs train against may be steering models wrong. A Contrary Research report argues that labs now train on the same tests used to rank them, so a flawed test pulls the whole climb in the wrong direction. Its main case is Meta’s Llama 4, which ranked second on Arena as an experimental variant and 32nd once the public release was retested.
- A new test grades AI on the office file it hands back, not its reply. Office Intelligence grades AI on the file it hands back, not its chat reply. In its finance case an agent must return an editable Excel workbook and leave unresolved differences open instead of papering over them. The page says two of its three parts were accepted at NeurIPS 2026, which is its own claim.
Cheques With Strings Attached: Money in Odd Shapes
One fund gets paid by how a marketing budget performs instead of by owning a piece of the company. A robotics startup is valued at $1.45 billion for a gentle grip.
- Otter raises $320 million to buy customers, and no shares are sold. Otter raised $320 million from General Catalyst’s fund, and no shares were sold: the fund takes its return from how the marketing performs. That is more than six times the $50 million Otter had raised in total before. TechCrunch notes the company gave no figures for monthly active users or churn.
- Robot startup Eka is valued at $1.45B for a hand that spares raspberries. Eka Robotics raised $150 million at a $1.45 billion valuation for a robot hand trained on touch in simulation, which one outside roboticist watched meet a raspberry very gently. The company says it succeeds on more than 95% of pick attempts. Every performance figure is Eka’s own, and no customer has it running in production.
Less Is More: Building Machines That Keep Running
Meta’s engineers explain what broke when one agent knew everything and what breaks when thousands of chips run as one. Amazon, meanwhile, wants developers to try its chips without rewriting a line.
- Meta’s AI agents got more reliable once the one in charge knew less. Meta’s first agent for loading data columns into its recommendation systems was given every rule and began mixing up subsystems. The fix was a router that does no domain work and hands each job to one of 14 specialists. Setup that took three to four days by hand now takes under an hour, though Meta gives no failure rate for the first version.
- Meta rebuilt an Ohio data centre to train its biggest AI models. Meta says its Prometheus site in New Albany, Ohio trained the Muse models, and that failures fell 99 percent in a year from a mid-2025 peak of nearly 1,000 events per 8,000 GPUs per day. It calls the cluster gigawatt-scale and gives no exact GPU count, and the post has no independent benchmark results.
- Amazon makes its AI chips run the code teams already write. AWS says its Trainium chips will run native PyTorch through TorchNeuron, an Apache 2.0 open source beta where moving a model to a device named neuron is the only code change. The performance evidence comes from partners: Deepgram reports about 60% more audio seconds per dollar than its GPU baselines.
Today’s Quick Hits
- Exa says its new search beats every test it ran, with no scores. Exa says its new search for AI programs beats every test it ran, but published no scores and named no rival. The checkable parts: a $4 Lite tier per thousand requests against $7 for standard, and an index of over 2 trillion addresses.
- Anthropic reportedly nears another $45 billion computing deal. Bloomberg reports Anthropic and Nscale may announce a second $45 billion computing purchase, about a month after the first. Nscale’s own investors are pushing back on its proposed $30 billion valuation as it pursues a stock market listing.
- Free plumbing, if the plumber films it for robots. AI lab HUMXN will pay for plumbing, HVAC and electrical repairs in three cities. The price is a worker wearing sensors that film the job inside your home to train robots. Tom’s Hardware reports no word on what happens to the footage.
- Google’s Notebook now shows you the plan before it builds your slides. Gemini Notebook now lets you edit an outline before it builds slides or a video overview, through a Review toggle that exposes the plan. Google says it came from user feedback and reaches every user worldwide over the coming week.
- Cloudflare previews a model that reads documents and returns the number. Cloudflare previewed Clairic, a model that pulls amounts, dates and estimates out of emails and receipts, each with a confidence score. Open weights are promised without a date, and its cost edge over a larger model rests on one test of its own.