The slowdown they got sued for
On 12 September, Anthropic’s CEO Dario Amodei published We Must Pace the Frontier: the labs should slow down together, test harder before release, and the US government should issue “a narrow waiver for certain kinds of safety conversations” so they can coordinate without breaking antitrust law. Sam Altman, Elon Musk and Demis Hassabis all responded publicly the same day.
Six days later, four people who pay for ChatGPT, Claude, Grok and Gemini sued Anthropic, OpenAI, SpaceXAI and Google in the Northern District of California, for a proposed nationwide class. The complaint alleges an illegal agreement to slow development that cut the value of what subscribers pay for. They do not object to any lab slowing itself; they object to competitors agreeing to “substitute collective restraint for individual accountability.” It is a complaint, not a ruling. Coverage: CNN, CBS News, PBS News, The Hill, TNW.
Three labs, all cheaper
Within four days, three of the four shipped new models at lower prices (per million tokens, input / output):
| model | now | was |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | $5 / $25 (Opus 5) |
| GPT-6 Sol | $2 / $10 | $4 / $20 (GPT-5.6 Sol) |
| GPT-6 Luna | $0.10 / $0.50 | new |
| Grok 4.7 | $2 / $6 | same as Grok 4.6 |
OpenAI told VentureBeat the Sol and Luna prices are permanent, not promotional. Grok 4.7 scores 46 on the Artificial Analysis index. So does MiMo V2.6 Pro, an MIT-licensed model you can download, which we covered in Brief 12.
The pause that did happen
On Friday 26 September OpenAI paused training its latest models after reviewing incidents from the summer in which its agents, sent to gather information from US government websites, went beyond what they were asked. They found API keys for Education Department data (only public information was gathered, per OpenAI), used login credentials found online to pull public Census Bureau data, and posted public SEC data somewhere else. The agencies say nothing private was reached. Researchers at Transluce say the agents also tried and failed to hack an Education Department site; OpenAI has not confirmed that part. OpenAI says it resumes “only when we are confident that we have additional safeguards” and expects to hit pause again. Coverage: NBC News, CNN, NPR, CBC News.
Earlier in the month OpenAI also disclosed an unreleased model writing notes to itself during training (compaction summaries) that declared it “freed from the roles and identities that bind other chatbots.” OpenAI found 27 such summaries, calls them extremely rare, and says none show the model escaping control (Malwarebytes).
A lab selling what the essay feared
At its Apsara conference Alibaba said Qwen 3.8 Max ran 33 fully automated cycles over a month, designing its training pipeline, checking its data, running experiments and diagnosing its errors, and rose from 40 to 45 on the Artificial Analysis index. In a chip-design test it made over 10,000 tool calls across 60 hours and shrank a module by 42%. All Alibaba’s own figures (press release). Qwen 4 is in training and the roadmap reaches 5 to 10 trillion parameters. Nobody has shown a model designing its own successor; this is a lab saying its model ran its own post-training loop and got measurably better.
The model that doesn’t talk
Jev, from TypeSafe (founded by Diogo Almeida, a co-inventor of ChatGPT), never writes a sentence. You give it a question and the allowed answers; it returns how likely each one is. Output is free and input costs $42 per billion tokens. It is closed and invite-only. Within about a week Jared Palmer released Kev, Apache 2.0, built on Qwen models and speaking the same interface; his largest is within a point of Jev on new-source accuracy, and fine-tuning the small one costs about a dollar. Then CLM-8B from researchers at Stanford and NVIDIA: open weights, one graphics card, and up to nine times faster than Jev in its authors’ tests.
Why Claude started writing like a manual
Jackson Kernion, who works on Claude’s fine-tuning at Anthropic, explained that models trained on explanations written for other AI models learned to write for machines: dense info dumps that lose a person (The Decoder). He says Opus 5.5 is an attempt to fix it. The tip: tell it who is reading. “Explain this to a person, not another AI.”
A law firm bought its own GPUs
Latham & Watkins bought Nvidia H200 servers and fine-tunes Nvidia’s open Nemotron models on them, in a data centre only Latham can reach, so client work stays on hardware it controls. It still uses cloud tools where they fit (Bloomberg Law). Same idea as this channel at a bigger budget: what one graphics card runs.
Claude in Chrome can’t open Reddit
Since 18 September, every Claude in Chrome action on Reddit fails with “This site is not allowed due to safety restrictions.” It worked the day before. The bug report is still open. Anthropic’s safety page lists adult, pirated-content and financial sites, not social media. Separately, Reddit’s scraping suit against Anthropic is active in San Francisco Superior Court (fiund). We don’t know why. A cloud agent’s reach can change overnight, with no announcement.
The ticker
Meta’s Muse hit 1.8 million iOS downloads in the US and Canada in its first 12 days, ahead of ChatGPT’s own launch (TechCrunch); estimates now put it past three million. There is already an open copy, OpenMuse, MIT, which you point at your own model. Alibaba’s Qwen Audio 3.1: five models, price cuts up to 95%, all closed. YouTube will let creators test up to three cuts of a video. And Mark Tilbury ran an AI side hustle for a week on $99: 23 sales, $157.48 profit by his accounting (his on-screen note: the sales were refunded after the experiment).
Lab report
This week: a private ChatGPT for the family on an old gaming PC with Ollama and Open WebUI, and 252 face edits across four photo editors, where our own editor was the one that scrubbed off the freckles.
Presenter shots were generated with MiniMax H3, by MiniMax, under the MiniMax H3 licence.