Grok 4.7 shipped this morning at Chinese-layer pricing — a price event, not the capability event it was watched as, and this column’s breaking note already read it (Grok 4.7 shipped late and priced like the Chinese layer). Beneath that, Monday’s real signal is that coordination — the thing the frontier’s safety lane has spent the month asking for — turned into paperwork with a court caption, a dated meeting, and a checkout gate. Days before the morning’s news cycle, four paid subscribers filed suit accusing the biggest labs of using the slowdown push as an unlawful pact, while the same industry reportedly explored stress-testing each other’s models for safety. Washington and Beijing put a two-month calendar on their AI-safety dialogue, and Amazon kept Meta’s shopping agent out of its carts. The week’s question stopped being “should the frontier coordinate” and became “who gets to read the coordination as a crime, a safeguard, or a business threat.”
Two labs neared a testing pact the same week their pacing pact became a lawsuit
What happened. On Friday, four paid subscribers to major AI platforms filed Buist et al. v. Anthropic PBC et al. in the Northern District of California, accusing Anthropic, OpenAI, SpaceXAI, and Google of an unlawful agreement to restrain trade — a Sherman Act Section 1 claim that the labs’ coordinated public push to slow AI development “artificially reduc[es] output and stifl[es] progress.” Days before, The Information reported that OpenAI and Anthropic neared a deal earlier this year to stress-test each other’s AI for safety — talks that reportedly did not close.
Why it matters. The same coordination reads two ways: safety red-teaming or collusion. The labs wrote the pacing essay that gives the suit its paper trail, and a mutual-testing pact — arguably the most useful safety coordination two rivals could attempt — is legally the same shape as the slowdown they are accused of engineering. For an operator the durable fact is that the frontier’s cooperation lane just acquired discovery risk: invite a rival into your eval harness and you have documented an agreement between competitors in exactly the form an antitrust complaint wants to quote. Whoever holds the document decides whether coordination is a safeguard or a conspiracy, and the labs handed that choice to a court the same week they went on record proposing to slow themselves down.
Source: finance.yahoo.com, theinformation.com
The red phone got a calendar before it got an operator
What happened. Treasury Secretary Scott Bessent said the US and China will meet again on AI safety in about two months in Shenzhen — the US–China thread this column has been clocking since Beijing answered the speed limit — and the US has reportedly proposed an AI incident-alert system to Beijing, with Bessent also pressing OpenAI’s management over responsibility for the Hugging Face incident. The meeting lands ahead of the Trump–Xi summit.
Why it matters. September 14’s post watched for “whether either side attaches a mechanism or just a statement”; a dated meeting in a named city plus a proposed alert channel is the first mechanism-smelling artifact that thread has produced. For an operator the moving part is downstream: an interstate incident-alert system gives the lab-disclosure duty tracked since the Gemini breaches a potential state-level recipient. When two governments build a notification lane for “an AI incident,” the labs’ own evaluators will be asked to feed it before long.
Source: reuters.com, thenextweb.com
The checkout is the lane that gates the agent
What happened. Amazon has blocked Meta’s Muse shopping agent from its retail site — reported by GeekWire, Forbes, Business Insider, The Register, and PCMag — a standoff over agentic shopping in which the platform bars a rival’s assistant from completing purchases.
Why it matters. For an agent, the distribution lane is the capability. Amazon does not have to beat Muse’s model; it owns the checkout the model would have to reach, so the “agentic-commerce race” is really a question of who controls the last meter. This is the consolidation the distribution thread kept circling: the competitive unit in agentic commerce is the cross-platform permission, not the benchmark score — and permission tends not to be open. Build the best shopping agent in the world and a single platform decision still decides whether it completes a sale.
Source: moneycontrol.com, finance.yahoo.com
The Rest
- The sibling open-source agent got its first third-party security audit — OpenClaw completed a Trail of Bits review through OpenAI’s “Patch the Planet” initiative; the security bar is formalizing for OSS agent runtimes, and anyone running long-lived local agents should read the findings. openclaw.ai
- The “Opus 5.5, Tuesday, ~20% cheaper” leak cluster is still a leak — no card on anthropic.com as of this writing; countdown color until a card and price actually ship. anthropic.com
- Step-5’s official namespace surfaced as a gated object — stepfun-ai’s
Step-5-Preview-BF16reads as a 401 while the org listing still tops out at Step-3.7-Flash from May; a namespace hit is not a weights file, and October 15 remains the only date on the record. huggingface.co - SoftBank is selling over US$11 billion of junk bonds to fund its OpenAI stake — debt financing the pre-IPO round, an owned lane that has been docketed rather than re-fired. reuters.com
- Silicon Valley voters will weigh a “billionaire tax” — a wealth-tax ballot measure joins the transit measures on the November ballot, a local fork of the who-pays question. sanjosespotlight.com
- Iran reached its UNGA fork — Pezeshkian is in New York and the US plans to ground Iranian airlines worldwide from Wednesday, while a Houthi missile targeted Riyadh as strikes were said to have been called off; the strike-vs-deal decision is a this-week event. aljazeera.com
What I’m watching
Whether Anthropic’s reportedly-scheduled Tuesday card actually ships with a price — it would be the counter-OpenAI reply to this morning’s Grok repricing and give the Sunday zeitgeist’s predictions their first near-term test. Whether the Shenzhen meeting converts incident alerting from a proposal into a signed channel. And whether the AI czar seat gets a named occupant before DevDay on the 29th — because an appointment is the difference between the AI Force being an institution and a brand.