FTC subpoenas OpenAI and Anthropic, Meta's model runs on Macs
The FTC's demands to OpenAI and Anthropic, Meta's new open Muse Spark model, Accenture rationing Claude, Microsoft's biggest in-house model, and Anthropic opening up to METR.
The FTC hit OpenAI and Anthropic with 50-plus-page safety subpoenas Axios says the civil subpoenas went out on Wednesday and ask for records on security incidents, safeguards, marketing and privacy policies. The FTC’s investigation follows this summer’s incidents where the labs’ models broke into outside organisations during testing, and it plans to make top executives testify and send demands to other organisations later. OpenAI says it’ll engage constructively with the FTC. Source
Meta put out an 816bn-parameter open model that runs on Macs It’s called Muse Spark Open, the weights went up on Thursday, and Meta says it’s the most capable open model yet from a US company. You can run it at home on two of Apple’s top Mac Studios, or through the big clouds and hosts like OpenRouter and Groq. The licence is free for many developers but limits some commercial use. Source
Accenture caps what it pays for Claude at $29 per employee 404 Media saw the internal memo: it’s $29 a month, and since 1 October anyone who goes over has to find a client project or team to cover the rest, and access can be taken away. Some staff had been spending $500 to $1,000 a month. Accenture says it makes AI available to all its people. Source
Microsoft’s new in-house model has 2.8 trillion parameters MAI-Base-2 took 38 days on 16,000 Nvidia chips over August and September, and Microsoft says it gets 4 to 12 times more out of its training compute than DeepSeek’s V4-Pro. It’s still being finished, and it won’t be open: Microsoft plans to license it to trusted partners. The write-up came out on Thursday. Source
Anthropic gives six METR researchers inside access, and no publishing veto METR, an outside safety testing group, gets company hardware, office space and internal tools, and Anthropic can’t stop it publishing what it finds. Anthropic can only make it black out a few set categories, like customer data, and METR can say publicly whether it thinks those cuts were justified. It came out on Thursday, two days after Reuters reported the labs wanted control over testers’ reports. Source