Four threads run through today’s AI news, and the first is who supplies the compute. OpenAI is close to a $30 billion round at $1.25 trillion, $150 billion under its ask, while its Jalapeño chip beat Nvidia’s Blackwell on performance per watt.
The second is agents getting everyday accounts. 1Password will let them log in without ever seeing your password, and Anthropic now runs Claude inside Google Cloud billing.
The third is small models doing odd jobs. Simon Edwardsson trained a 360-million-parameter model to imitate a terminal, and Perplexity has one model play both teacher and student.
The fourth is oversight. METR and Redwood Research spent three weeks inside OpenAI testing how its agents are watched, and nine Claude models now carry a text watermark.
Money and Silicon: A Lower Price Tag and Chips Built to Waste Less
OpenAI’s raise came in under the number it first asked for, and two hardware stories show how much the company and its rivals now care about what a chip delivers per watt.
- OpenAI is close to a $30 billion round, priced $150 billion below its ask. CNBC’s sources say UAE investors lead and OpenAI hopes to have signatures before next week ends, at a $1.25 trillion valuation against the $1.4 trillion Bloomberg reported. CNBC says the size is still up to $30 billion, and nothing is signed.
- OpenAI’s first inference chip beat Nvidia’s Blackwell on power efficiency. Hardware chief Richard Ho credits high utilization, not bigger peak specs, for the lead in SemiAnalysis InferenceX tests on open models. Neither the interview nor Tae Kim’s newsletter gives the size of the lead, and Kim says Nvidia’s Vera Rubin could change the rankings.
- Cerebras says it runs a 3-trillion-parameter model across 59 chip wafers. Each wafer holds 44 GB, so Cerebras passes Kimi K3’s layers down a line of wafers and claims 1,000 to 2,000+ tokens per second. The figure is the company’s own, and the post gives no number for tokens per second per megawatt, the yardstick it says matters most.
Agents Get Everyday Accounts: Logins and Cloud Invoices
Before agents can run errands for you, they need a way to sign in and a way to be paid for. Two announcements this week go after exactly that.
- 1Password lets AI agents sign in to sites without ever holding your password. The agent asks, you approve inside 1Password, and it fills the login itself, at no extra cost on Individual and Family plans on 1Password.com. Only Meta’s Muse and Grok Bot are due within days, while Claude Duet and ChatGPT Work are listed as coming soon.
- Anthropic now sells Claude through Google Cloud, on one Google invoice. Customers sign in with their Google credentials and can count Claude spend against commitments they already hold, with new features arriving on launch day. Anthropic acts as the data processor, and the post does not say whether per-token prices match its own platform.
Small Models, Odd Tricks: A Fake Terminal and a Self-Teaching AI
One solo builder shrank a computer terminal into a model that fits in a browser tab. One company says a single model can coach itself better than a bigger teacher can.
- A 360-million-parameter model imitates a computer terminal key press by key press. Simon Edwardsson’s WorldTUI got 80 of 106 test commands right, against 52 for the original Qwen model when it was only shown examples. Chaining familiar commands in new ways still breaks it, and every number comes from his own tests.
- Perplexity trains one model to play both teacher and student. Perplexity says its JOLT+ method beat a standard reinforcement learning recipe on five math, coding and agent setups, and that a fitted curve on one math test needs 6.5 times fewer training tokens. The code is not out yet, so nobody outside the company has rerun it.
Watching the Machines: An Audit, a Watermark and a Treaty Sketch
Three different attempts to keep track of what AI is doing, from inside one lab, across every Claude output and between two governments.
- Outside testers found ways OpenAI’s coding agents could dodge its monitor. METR and Redwood Research spent three weeks inside OpenAI and found gaps, including compaction events that left roughly 10 percent of eligible Codex traffic unwatched from May to July. OpenAI says every production vulnerability is patched, and METR says it has not checked that.
- Nine Claude models now carry a text watermark that does not identify the user. O’Reilly’s Bruce Hopkins says Claude Desktop, Claude Code and the API all emit the mark, built on Google’s SynthID-Text. The only published accuracy figure in his piece, 95 percent at 400 tokens, is OpenAI’s own for its rival mark.
- A LessWrong post sketches how the US and China could verify an AI slowdown. The author puts Claude Opus 5.5 at 167 on the Epoch Capabilities Index and China’s Kimi K3 at 157, roughly a year of progress, and says a US-only pause would hand China the lead. The checks run from satellites to inspectors to chips that report on themselves, and no such chip exists yet.
Today’s Quick Hits
- NVIDIA tool makes a video and matching 3D animation from one text prompt. NVIDIA Research’s Joints, Camera, Action! writes a video and matching motion data for any 3D character skeleton from one prompt, trained on nearly 200,000 pairs. NVIDIA shows clips against three rival tools but publishes no scores.
- Pichai explains why Google’s Gemini 3.5 Pro never shipped. Sundar Pichai said, in an exchange Alex Heath posted on X, that Google put its coding fixes into the smaller Flash model first because it allows more training iterations, and by the time they worked, Pro’s base training was already done.
- Anthropic hands Agent Skills to the Agentic AI Foundation. Anthropic donated its format for teaching agents specific jobs to a body under the Linux Foundation. More than 70 platforms support it, and developers have published 1 million skills with nearly 280 million installs.
- Shattered explains what sits under Microsoft’s Decision-1 model. Shattered’s explainer says Decision-1 sits on the 9-billion-parameter Qwen3.5 base, takes text only, accepts up to 32,000 tokens and returns a probability for each preset answer. It is a third-party write-up, not Microsoft documentation.
- One writer says Flapping Airplanes’ $5 billion price is a weak clue about AI timelines. I.M.J. McInnis argues the reported jump from $1.5 billion to $5 billion says little about how soon powerful AI arrives, and discloses that Lux Capital, named in the talks, invested in his own company.