Daily Brief · 21 September 2026
Subscribers sue OpenAI, Anthropic, Google, and xAI over alleged AI slowdown pact
Four paid users of ChatGPT, Claude, Gemini, and Grok filed a proposed class action in the Northern District of California accusing OpenAI, Anthropic, Google, and SpaceXAI/xAI of coordinating to decelerate frontier development under a safety banner. The complaint points to Dario Amodei’s 12 September “pace the frontier” essay and same-day public agreement from Sam Altman, Elon Musk, and Demis Hassabis, plus a July lab-staff statement urging government support for a global slowdown. Plaintiffs say any such pact would cheapen paid subscriptions; the companies have not answered in court. The case tests whether CEO alignment on safety can be cast as an antitrust agreement—and lands as Washington still resists industry calls for a waiver.
What else matters
Shanghai AI Lab ships Atria Dawn Preview weights under MIT
Shanghai Artificial Intelligence Laboratory released Atria Dawn Preview, an agentic instruct model on the 744B-parameter MoE GLM-5.2 base with a 256K context. Weights (BF16 and FP8) are on Hugging Face and ModelScope under MIT; the lab pitches a full loop of plan → code → experiment → error recovery baked into the weights, with an API that rejects images and PDFs. Self-hosting is multi-accelerator territory—BF16 alone is about 1.5 TB on disk.
Hugging Face — internlm/Atria-Dawn-Preview
StepFun opens Step 5 Preview API: 600B MoE at $1 per million input
StepFun announced Step 5 Preview on 20 September: a 600B sparse MoE with 27B active parameters, a 1M-token context, and API pricing of $1 / $2.70 per million input/output tokens plus a 95% cache discount. Artificial Analysis scores it 44 on its Intelligence Index. The API and Studio opened the same day; full open weights are scheduled for 15 October—the Hugging Face repo is still essentially empty.
TypeSafe AI launches Jev, a sub-second decision model—not a chatbot
TypeSafe AI, founded by former OpenAI RLHF researcher Diogo Almeida and freshly backed with $40 million, unveiled Jev: a text-and-structured-data model that returns decisions in roughly 70–500 ms at $0.042 per million input tokens (output free). There is no chat UI and no generative prose—Almeida’s bet is that automation needs fast choices, not conversation. Early-access benchmarks are TypeSafe’s own agreement scores against frontier models, so treat the headline speed/cost multiples cautiously.
xAI puts Grok 4.6 on the API with 500K context
xAI’s developer release notes now list Grok 4.6 on the xAI API: text and image input, text-only output, a 500K context window, and selectable reasoning effort (low / medium / high default / xhigh). Pricing is $2 / $0.50 / $6 per million tokens (input / cached / output) below 200K prompt tokens, doubling above that band. Separate from earlier Copilot availability, this is the developer-metered path.
Worth your time
Canada and Germany’s joint LawZero announcement (CAD 150M + EUR 100M for Bengio’s Scientist AI)—the official newswire, useful context beside the slowdown lawsuit.
