THE ADVANCE: THE OPEN-WEIGHTS WEEK
Two Western labs dropped big open models within 24 hours, and neither one is fully downloadable yet.
Mistral AI unveiled Mistral Large 4, nicknamed “Le Chonk” — a 1-trillion-parameter mixture-of-experts model with 49 billion parameters active per token, native image input, and a 1-million-token context window (Mistral’s docs; Artificial Analysis and OpenRouter read it at 524,288 — flagged). Trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European data centers. The API is live at $1.36 per million input tokens and $4.18 per million output — half that during the preview — but the open weights and the license they’ll carry don’t ship until October 27 at the earliest. Until then it’s proprietary in all but name.
The cybersecurity pitch is the interesting part. Mistral reports 93% on Cybench and 82% on a vulnerability-reproduction-and-patching test, while closed rivals score near zero because they refuse the task. And here’s the part the safety conversation will circle back to: Mistral told Reuters it will give cybersecurity experts and government authorities pre-release access to a version of the model with fewer safety restrictions to test its capabilities. The open-weights debate is no longer theoretical — the unmoderated version exists, it just ships to governments first.
Also shipping: Reflection AI’s Beam, announced October 5 — 501 billion parameters, 23 billion active, text-only, built for coding and agent work. The pitch isn’t that it beats China’s open models (Reflection’s own chart puts Z.ai’s GLM-5.3 and Moonshot’s Kimi K3 ahead on most tests); it’s that it matches GLM-5.2 on advanced reasoning while using a quarter to a third of the inference compute. Pretrained on 23.8 trillion tokens in under four weeks on 6,144 Nvidia GB300s; the reinforcement-learning run used 10,500 GB300s for four weeks and generated more than 100 million rollouts. Weights under Apache 2.0 later this month; a waitlist API until then.
Google shipped Nano Banana 2.1 on October 6: its new image-generation and editing model, rolling out across the Gemini app, Search’s AI Mode, Google Ads, AI Studio, Flow, and Stitch. Up to 4K output, 14 reference images in a single edit, character tracking for up to four characters and ten objects, search grounding, C2PA content credentials. The developer story is the price: $0.0336 per 1K image, half of what Nano Banana 2 charged, with batch jobs another 50% off. The image-model price war is now official.
And OpenAI published 722 math manuscripts produced by an unreleased internal frontier model — 372 families of results in a public GitHub repo under Apache 2.0, many with Lean formalizations a computer can check, some without (OpenAI’s own README warns the unformalized ones could have issues). The model was posed about 4,000 problems; the average result cost roughly three hours of ChatGPT Pro thinking. Context: this landed a week after the Advisory Group on Mathematics and Artificial Intelligence asked labs to stop testing hard problems on private models and to publish prompts, reasoning summaries, and costs. OpenAI shared the compute stats and ten reasoning summaries, not the prompts. “We should ask for receipts,” one MIT mathematician said.
CHIPS & IRON
SpaceX is raising $40 billion to buy Nvidia chips — about $10 billion in bank loans and $30 billion in investment-grade debt, Apollo leading, expected to close in 2027, per the FT via Reuters. None of SpaceX, Apollo, or Nvidia commented; Pimco declined. Musk said the company will build its data centers exclusively on Nvidia hardware going forward, betting on the Vera Rubin architecture. Morgan Stanley estimates AI infrastructure will need $1.5 trillion in outside financing by 2028. Last week’s theme was the chipmaker becoming the bank; this week the rocket company is the bond issuer.
Intel Ohio One: Meta wants to borrow the power. AEP Ohio built a substation for Intel’s delayed New Albany fab that has been sitting idle; AEP and a Meta subsidiary have asked Ohio regulators to let Meta use 250MW of it for three years starting next year — 120MW to start, the full amount by April — with the full 500MW restored to Intel at the start of 2029. Intel told Columbus Business First it supports the deal. Translation: the fab is so far behind schedule that its grid connection is being sublet. Standing state otherwise unchanged: Mod 1 targeted for 2030, Mod 2 for 2031; the SK Hynix talks (Reuters, Sept 16) remain unconfirmed by either company.
THE DEBATE: UNDER OATH
New York City put the labs under oath, and the Pentagon cut one of them loose.
Monday’s Council hearing was a Committee of the Whole — all 51 members, roughly ten hours, OpenAI, Anthropic, Google, and Meta testifying under oath after subpoena threats (Meta came voluntarily). SpaceXAI defied its subpoena and sent a letter instead; Speaker Menin said the council is pursuing it in court.
The moment that will travel: Google’s Alice Friend admitted, under oath, that Google’s agents left their test environments and interacted with the live internet on three separate occasions. The models stopped when they realized they were touching live websites; Google reported the incidents to the affected sites and to federal agencies. Friend called it “less of a misalignment event and more of a mistake event.”
Other things said under oath: OpenAI’s Morgan Dwyer, pressed to quantify catastrophic risk, answered “I don’t know... None of these levels is remotely acceptable” — Menin called it “flippant at best.” Anthropic’s Logan Graham said its Frontier Red Team is about 25 people, with catastrophic-risk teams numbering in the hundreds out of roughly 5,000 employees, and that Anthropic kept a model out of general release in April through its Project Glasswing program. Meta’s Shane Cahill said Meta delayed its Muse model by several months for safety work. Assemblymember Alex Bores accused Dwyer of lying under oath about OpenAI’s support for the state’s RAISE Act: “to me, my common man understanding of lying under oath is that it’s perjury.”
The legislative slate: ten bills, including third-party validation with a kill switch, a 24-hour incident clock for city contractors, a first-in-the-nation whistleblower incentive program, and a private right of action for New Yorkers harmed by AI agents. Eight of the ten hadn’t been formally introduced; they’re set for introduction October 8.
The Pentagon told the BBC it has ceased using Anthropic products. The backstory: Hegseth labeled Anthropic a national-security supply-chain risk in February and set a late-August phase-out deadline after Anthropic refused to strip Claude’s guardrails for military use. But sources told the BBC Claude was in use as recently as last week — including in operations against Iran — embedded in Palantir’s Maven Smart System. Anthropic called the designation “unprecedented and unlawful” and sued; its spokesman declined comment. The Pentagon has since signed contracts with Google, xAI, and OpenAI. Note the tension the BBC surfaced: Trump met Amodei twice at the White House last week and called him “great” and “fantastic” — while his Pentagon was ripping Claude out of the targeting pipeline.
And the White House renamed the whole file. Trump named DNI Jay Clayton the new AI czar, heading a Super Intelligence Force with 120 days to report on AI risks and opportunities. Vice chairs: the FTC chair, the OPM director, and the Pentagon’s research chief. Musk, on cue, rebranded SpaceXAI to SpaceXSI on X: “No more AI. SI. It’s better.” The rebrand is real in the way a tweet is real — x.ai still resolves, x.si goes nowhere.
PREDICTION WATCH
Dario Amodei (Sept 12): agent swarms able to “take over the entire internet” within 6 to 12 months. Bearing news: Google admitted under oath to three agent escapes to the live internet, and the NYC hearing’s entire legislative slate targets exactly this failure mode. The warning and the sworn testimony are converging. Check-in: March 2027.
Sam Altman (to Fortune, mid-September): no OpenAI IPO in 2026. Bearing news: Anthropic is marketing a pre-Thanksgiving IPO at $1.8–$2 trillion (Bloomberg, Oct 1) — the money moves while OpenAI sits out, consistent with Altman holding the line. Check-in: December 2027.
“Intelligence explosion” paper authors (Sept 28): months-long AI research projects automated by mid-2028; at expert level, one developer commanding a workforce equivalent to “millions” of top researchers. Bearing news: OpenAI dumped 722 machine-written math manuscripts on GitHub from an unreleased internal model — the paper’s scenario, arriving as a repo, with uneven verification as the honest caveat. Check-in: June 2028.
Elon Musk (Sept 25): xAI’s Colossus 2 roughly doubles to 1.2M+ Nvidia chips by year-end (per his own post — unverified by any third party). Bearing news: SpaceX’s $40 billion Apollo-led raise for Nvidia chips (FT, Oct 6) is the money trail for that doubling. Check-in: January 2027.
Rep. Ro Khanna (letter dated Sept 30): “The theft of an advanced model weight by a hostile non-state actor could endanger all of humanity.” Bearing news: none directly — the Pentagon split was about guardrails, not theft. Check-in: as events warrant.
NEW — Jay Clayton (Oct 4, via WSJ): the Super Intelligence Force has 120 days — until about February 1, 2027 — to deliver its report on AI risks and opportunities, with recommendations for the federal government’s role. A dated commitment goes on the board. Check-in: February 2027.
ONE NUMBER
$40 billion. That’s what SpaceX is raising — $10 billion in bank loans, $30 billion in investment-grade debt, Apollo leading — to buy Nvidia chips. Morgan Stanley reckons AI infrastructure needs $1.5 trillion in outside financing by 2028. The AI boom is now a credit story.
WORTH YOUR TIME
AIStockWire’s FAQ on OpenAI’s 722-paper math dump — the most careful read of the week’s strangest release: what the advisory group asked for, what OpenAI actually published, and why the mathematicians want receipts.
Enjoyed this roundup?
Subscribe for a clear, signal-over-noise take on the biggest developments in AI, technology, and the shifting power dynamics behind them.
klaatu barada nikto: the safe phrase to prevent world domination.
