AI News: September 21, 2026

1. A Proposed Class Action Says the Safety Slowdown Was an Illegal Agreement

Anthropic, OpenAI, SpaceXAI, and Google were named in an antitrust suit filed Friday in the U.S. District Court for the Northern District of California, on behalf of four plaintiffs who pay for ChatGPT, Claude, Grok, and Gemini and seek to represent a nationwide class of paid subscribers. The complaint argues that coordinating a deceleration of frontier development is itself an anticompetitive act, because subscribers get less for the same money than unconstrained competition would produce.

The alleged coordination point is September 12, when Anthropic CEO Dario Amodei published an essay urging industrywide cooperation on slowing development in favor of safety work, and Sam Altman, Elon Musk, and Demis Hassabis each publicly endorsed it the same day. The filing also reaches back to a July 2026 statement signed by senior staff across several labs acknowledging “intense competitive pressure not to unilaterally slow.” Amodei’s original essay had anticipated exactly this exposure, asking for government mediation or a narrow waiver covering cross-lab safety conversations. None of the four companies responded to requests for comment. Source

2. StepFun Opened API Access to a 600B Sparse MoE and Dated the Weights for October 15

StepFun announced Step 5 Preview, a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active per token, a 1M-token context window, and multimodal input. The company positions it for agentic work, citing software engineering and professional knowledge tasks with particular strength in finance. API and Studio access opened the same day; the full weights are scheduled for release on October 15, so independent deployment and self-hosted performance numbers are not yet verifiable.

Pricing is $1 per million input tokens and $2.70 per million output tokens, with a 95 percent cache discount bringing cached input to roughly $0.05 per million. Artificial Analysis places it at 44 on its Intelligence Index, level with Kimi K3 Max and at a fraction of GPT-5.6 Sol’s price. Source

3. A Researcher Traced ChatGPT’s Ad Pixel Binding Off-Site Browsing to Accounts

An independent researcher publishing as Buchodi documented how OpenAI’s advertising measurement pixel links browsing on third-party sites back to ChatGPT accounts. The mechanism is an __obi cookie set on .openai.com with a one-year lifespan, signed as a JWT that binds the identifier to a user account. Advertisers install OpenAI’s pixel SDK the same way they install Meta or Google tags, and subsequent requests to bzr.openai.com carry the cookie alongside page context such as products viewed and articles read.

The measurement covered 12 commercial sites including Chewy and Wayfair across 12 distinct pixel IDs. In the captured traffic, identity fields scraped from the page outnumbered identity deliberately supplied by the advertiser 685 events to 255, covering email addresses, phone numbers, names, and location; some fields are hashed before transmission while postal codes travel unencrypted. The author’s own framing is that the adtech is unremarkable and the surface is not, given what users disclose to a chat product versus a social feed. The writeup reached 696 points on Hacker News. Source

4. Qwen-Image-2.1 Fits Generation and Editing on a Consumer GPU

Alibaba’s Qwen team released Qwen-Image-2.1, an open-weight image model at 7 billion parameters that generates and edits on high-end consumer GPUs. It supports transparency and accepts up to ten reference images in a single call, which matters for consistent-character and product-shot workflows that otherwise require fine-tuning. The team claims it beats closed models on image generation at that parameter count.

The license is the constraint: the released weights carry a research license that bars commercial use, and commercial deployment requires a separate Qwen license. Source

5. Alibaba Open Sourced a Code Review CLI That Keeps the Deterministic Parts Deterministic

Alibaba open-sourced OpenCodeReview, a command-line code review tool that splits the job between deterministic pipelines and an LLM. File selection, bundling, and rule matching run as ordinary deterministic stages, and the model is invoked for the review judgment itself rather than for deciding what to look at. That split is the design point: it makes the expensive and non-reproducible step a smaller fraction of the run, and it keeps scope selection auditable. Source

6. Tencent Split a Model That Talks From a Model That Works

Tencent detailed Gander, a system that processes speech, images, and text while running tasks in the background. The architecture separates a “cerebellum” that sustains the conversation from a swappable “brain” that searches files, writes code, or handles other long-running work, so the user can interrupt or redirect mid-task without waiting for the working model to finish.

The tradeoff shows up in the benchmarks. Gander interrupted users in 8 percent of cases, less often than competing systems, but it trailed those systems on task accuracy. Source

7. Runway Wants Video Generation to Be a Stream You Steer, Not a Render You Wait For

Runway is pushing AI video toward real-time streaming, generating frames as the user prompts rather than returning a finished clip after a wait. The approach builds on GWM-1, the company’s world model, which already generates video frame by frame. Runway frames the applications beyond creative tooling, pointing at robotics and autonomous driving, where a controllable frame-by-frame world model is closer to a simulator than to an editing tool. Source

8. Simulated Students That Fail Realistically Produced Better Tutors Than GPT-5.4

Researchers at Microsoft and the University of Illinois built StudentSim, which reconstructs an individual student’s error patterns from limited data so an AI tutor can get fast, cheap feedback without a human in the loop. Across 60 students in chess, English, and math, tutors trained against StudentSim outperformed GPT-5.4, and a chess tutor trained this way drew the highest expert ratings among three versions tested. The mechanism that matters is fidelity of failure: a simulated learner that makes plausible mistakes gives a tutor a usable gradient, where one that answers correctly or randomly does not. Source

9. Near-Daily AI Use in the US Went From 8 Percent to 19 Percent in Five Months

Surveys from Epoch AI and Ipsos put the share of US adults using AI almost daily at 19 percent in August 2026, up from 8 percent in March. The more telling number is the other side of the ledger: weekly-only use fell from 17 percent to 10 percent over the same period, which suggests the growth is existing occasional users converting to habitual ones rather than purely net-new adoption. Source

10. Anthropic Is Reportedly Pushing Its IPO From October to November

Anthropic is said to be delaying its public listing by a month to get strong Q3 results in front of investors first, following OpenAI’s own postponement. Reported investor expectations sit around a $2 trillion valuation. The complications cited are on the cost side and the risk side: infrastructure spend that includes roughly $1.25 billion a month for the SpaceX deal alone, plus unresolved security exposure that has not been cleared ahead of a listing. Source

11. Equity Asked Whether Slowdown Talk Survives Contact With the Business

TechCrunch’s Equity podcast took up whether the executives calling for deceleration are serious about it, arriving the same weekend the antitrust complaint in item 1 argued that they were serious enough to be actionable. The two readings are not compatible, and which one holds will be tested in the Northern District of California rather than in public statements. Source

12. Vocci Put Meeting Capture in a $249 Ring

Vocci launched a lightweight ring that records and transcribes meetings, at $249, adding a new form factor to a category currently dominated by pendants and phone apps. TechCrunch notes the obvious consequence of moving always-available capture onto jewelry: it is far less legible to the other people in the room than a phone on the table, which is a consent problem before it is a product problem. Source

13. ScrollEd Reformats Textbooks Into a Scrollable Feed

ScrollEd, a Palo Alto startup founded by student co-founders and spouses Utsav Gupta and Rebecca Neff, converts textbook material into an Instagram-style scrollable feed mixing video, audio, and quizzes. The company is pitching at TechCrunch Disrupt. Source