AI Safety Body & Anthropic Nasdaq IPO: AI News Sep 15, 2026

One Big Thing: the "pace the frontier" essay just became an institution

The headline today isn't a model. It's a rulebook. The Information reported September 14 that working groups below the CEO level at Anthropic, OpenAI and Google have met regularly since July to build a voluntary AI safety standards body, modeled on FINRA, the financial-industry self-regulator. Demis Hassabis floated a US "Frontier AI Standards Body" back in July; Sam Altman backed the concept on September 12, saying labs have to act without waiting for government; Dario Amodei's "We Must Pace the Frontier" essay the same day lines up behind it. The proposed body would set shared protocols for pre-release safety reviews, independent evaluations, and standardized risk assessments.

Watch for a joint press release that formalizes it, and for Congressional reaction before the fall regulatory session opens.

Microsoft got there first in writing. Satya Nadella said September 13 the company would publish a Code of Conduct for its first-party MAI models for public comment; the document dropped September 14 across three pillars—broad access at every layer of the stack, enterprise control of learning loops and models, and the behavioral spec itself. Nadella endorsed deliberate pacing and called for independent evaluators for superintelligence. Microsoft is the first hyperscaler to put its own model family's behavior out for open comment.

Here's my problem with the celebration. A body the frontier labs write themselves, to police the frontier labs, is not the same thing as oversight. The people drafting the access rules are the people being evaluated. And the timing is the tell: this coordination lands in the same week each of them closed a nine-figure compute commitment. Anthropic signed a $13.7 billion compute deal with Rum Group. Zhipu just raised $5 billion explicitly to fund "self-evolution." OpenAI confirmed it won't IPO in 2026—not because it's cautious, but to keep its burn-and-narrative freedom private. Self-regulation by the incumbents is a moat, not a brake.

Quick scan

• Anthropic IPO firms up. It has picked Nasdaq as the listing venue and expects to report a second straight profitable quarter. NVIDIA is still in talks to anchor the deal at up to $10 billion; Anthropic is seeking as much as $100 billion at a ~$2 trillion valuation, targeting a close before the November midterms. Its revenue run-rate reportedly passed $65 billion by end of July, up ~14x. (subject to vendor confirmation)
• Claude Code limit flip. The six-month 50% weekly-limit promo ended September 14 and was replaced by a permanent 25% bump over the pre-May baseline. Against the promo ceiling that's a 17% cut (125 vs 150 on a 100-unit base); Anthropic calls it a raise, BleepingComputer calls it a cut—both are arithmetically true. Separately, Claude Enterprise Smart Reports entered beta and the Admin API left beta.
• China keeps shipping. Zhipu closed ~$5B (Sept 13): $2B share placement plus $3B zero-coupon convertible, earmarked for next-gen GLM and a "fully self-trained" stack. DeepSeek V4.1-Flash is now live on Alibaba's Qianwen platform (Sept 14)—552B MoE, 8B/16B active, 1M context, MIT weights, native vision, cached input around $0.003 per million tokens; it reclaims the open-model cost crown. Alibaba's Ling-3.0-flash-VL hit OpenRouter with two weeks free plus open-sourced FP4/INT4 quants. ByteDance shipped the consumer Doubao Mobile Assistant with a physical AI key; the NaviX Ultra "AI agent phone" lands September 16. NVIDIA open-sourced Nemotron 3 Ultra's natural-language IMO 2026 pipeline (30/42, gold) with checkpoints and a 200-problem bench.
• Production Astra. OpenAI said September 14 that Perplexity deployed GPT-6 Astra across its end-to-end systems—docs, codebase edits, live monitoring—with far less human review than before.
• Funding & competition. Cohere is negotiating a $2–3B round at ~$20B (Canadian government involved, Schwarz Group leading). Kimi K2.8 Preview rolled out fully with 1M context to all members. The Google "rsi-model-liverl-le" API sighting from last week is still unexplained—Brin's recursive-self-improvement bet remains the thread I'm not dropping.

Editor's Take

I'll make a bet and eat it here: within six months this standards body produces exactly one durable artifact—a joint press release—and changes zero frontier releases. The danger isn't that it's toothless. The danger is the opposite: a credible industry body is the perfect political cover for Congress to stand down, which is exactly what the Senate "duty of care" bill needs to survive. Self-regulation that pre-empts legislation is worse than no regulation, because it looks like progress while locking the current ranking in place. What I'm watching: whether the body's first "independent evaluation" ever delays a launch, and whether Google—the lab with the unexplained RSI model—signs the same rulebook it helped draft.