three stories about the gap between what AI companies promise and what actually

1) Four AI Giants Are Sued Over an Alleged Pact to Slow Down

On September 18, four paying subscribers to ChatGPT, Claude, Grok, and Gemini filed a proposed class-action antitrust lawsuit in the U.S. District Court for the Northern District of California against OpenAI, Anthropic, Google, and xAI. The suit centers on September 12, when Anthropic CEO Dario Amodei published an essay urging the industry to collectively "pace the frontier" -- to slow deployment in favor of shared safety testing rather than racing each other to ship first. The complaint alleges that OpenAI's Sam Altman, xAI's Elon Musk, and Google DeepMind's Demis Hassabis each responded publicly in agreement the same day, and argues that this coordinated response is what crosses the legal line: the plaintiffs say they don't object to any one company individually deciding to move more cautiously, only to competitors agreeing with each other to do it together, which they call an illegal "shortcut" that substitutes collective restraint for individual accountability.

The concept, simply: antitrust law generally allows a company to unilaterally decide to slow down, raise prices, or hold back a product -- that's just a business choice. What it forbids is competitors agreeing with each other to do the same thing, because coordinated restraint among rivals can function like a cartel controlling supply, even when the stated reason is safety rather than profit. The lawsuit's whole theory rests on treating Amodei's essay and the same-day public replies as evidence of an agreement, not four separate coincidences.

Why it matters: this lands the same week the federal government signaled, through Trump's proposed "AI Force," that it has no interest in binding pre-deployment AI rules (Day 177). A private antitrust lawsuit is now trying to use existing competition law to force exactly the kind of external check that federal regulation isn't providing -- and however it's resolved, it will shape whether AI labs can ever again coordinate publicly on safety without having to prove, in court, that each decision was really made alone.

2) Meta's Muse Hid How Its Bookings Actually Work

Meta launched its personal AI agent Muse on September 8, and by September 18 it had become the #1 free app on the U.S. App Store, ahead of ChatGPT, Gemini, and Claude (Day 177) -- and its own employees found the cracks in it first. This week, reporting filled in a detail Meta didn't disclose at launch: Muse's travel booking runs on two entirely different systems depending on what you're booking. Flight search and booking go through a direct API connection to Duffel, a licensed travel-booking platform that stands behind the transaction. Hotel and car searches, by contrast, run through a browser that Muse operates itself on the open web -- clicking through ordinary hotel and rental sites the way a person would, with no equivalent backstop if something goes wrong. Separately, internal employee testing surfaced instances of private data exposure and unreliable operation before launch. Meta's approval-gate design, which requires explicit sign-off before Muse spends money or sends a message, catches the most obvious failure mode -- an agent acting without permission -- but does nothing to protect a user who approves a booking the agent already misread.

The concept, simply: an AI agent using a direct, licensed API is accountable in a way a browser-driving agent isn't -- the API provider has agreed to honor the booking and has support channels if it fails. An agent that instead clicks around the open web is doing what you'd do yourself, with all the same risk of misreading a page or booking the wrong dates, and no company standing behind the outcome beyond whatever that individual hotel site offers.

Why it matters: this is exactly the consumer-protection question this journal flagged as coming in yesterday's closing note -- and it arrives at a scale that matters, since Muse is now installed on far more phones than it was during Meta's internal testing. Which specific action an agent is about to take -- API call or open browsing -- determines who's actually responsible when it goes wrong, and right now users approving a Muse booking have no way to tell which kind they're approving.

3) StepFun Ships a 600B-Parameter Model at $1 per Million Tokens

On September 20, Chinese AI lab StepFun launched Step 5 Preview, a sparse Mixture-of-Experts model with about 600 billion total parameters and roughly 27 billion active per token, built on a 92-layer narrow-deep transformer. It's aimed squarely at long-horizon agentic work -- coding, software engineering, financial analysis -- with a one-million-token context window and text-plus-image input. API access is live now at roughly $1 per million input tokens and $2.7 per million output tokens, and StepFun says full model weights will open on October 15. On Artificial Analysis's intelligence index, Step 5 Preview scores around 44, which the company positions among the leading open-weight models available today. The concept, simply: a "sparse Mixture-of-Experts" model is built from many specialized sub-networks, but only routes each token through a small subset of them -- so a model with 600 billion total parameters can run at roughly the cost and speed of a 27-billion-parameter one, because that's all that actually activates per token. That's how StepFun can price a frontier-scale model at a fraction of what its total size would suggest. Why it matters: this keeps widening the gap this journal has been tracking between how Western and Chinese labs are handling their most capable models. The same week OpenAI kept its most dangerous GPT-6 Astra capabilities behind a vetted-access program (Day 175) and four U.S. labs are being sued over whether they can even coordinate on going slower, StepFun is doing the opposite: shipping a frontier-class agentic model at cut-rate pricing today, with the weights themselves set to become fully public next month.