Three stories worth your coffee this Friday. A pack of red-state attorneys general decided not to wait on Congress and went after OpenAI in their own courts. Meta dropped a giant open-weight model built to run on a couple of Macs. And Anthropic just handed an outside watchdog a company laptop and the keys to the building.
01 of 03
Eight states take OpenAI to their own courthouses
On Wednesday, Iowa Attorney General Brenna Bird stood up at the Iowa State Capitol with three other AGs: Kris Kobach of Kansas, Mike Hilgers of Nebraska, and JB McCuskey of West Virginia. They announced they are suing OpenAI, each one in their own state court. Arkansas, Arizona, Idaho, and Montana are also suing OpenAI in their own state courts, which brings the count to eight states.
The Iowa petition is already filed in Iowa district court. It says OpenAI failed to put proper warnings or usage limits on ChatGPT, especially for minors, and it points to false medical advice and responses that encouraged violence or self-harm. It leans on existing parent lawsuits, including one where parents say ChatGPT told their teenage son that combining drugs was safe, and he died. The legal hook is each state's consumer fraud act. Arkansas, Arizona, Idaho, and Montana are focused on what they call misleading statements OpenAI made about ChatGPT's safety.
The strategy is the interesting part. Hilgers said flat out he wants Nebraska parents sitting on a Nebraska jury deciding the harm. Kobach pointed out that more than 150 AI bills have been filed in Congress and most never got a committee vote. Bird also quoted Sam Altman telling Politico that "the world should accept some bad things happening," and said "we refuse to accept that."
Here is where I land. Splitting this into separate state cases is smart, and kind of brutal for OpenAI. One big federal case you can fight in one place with one set of lawyers. Eight state juries means eight different chances to lose in front of folks who are parents first and tech fans second. And Altman handing them that quote was a gift. Whatever you think about the merits, OpenAI is about to learn that "move fast" does not play the same in Des Moines as it does in San Francisco.
Source: Iowa AG petition (PDF) and Iowa Capital Dispatch.
02 of 03
Meta's Muse Spark Open is huge, open, and built for a pair of Macs
Meta released Muse Spark Open on Thursday and called it "the most capable US open-weight model to date." The numbers: 816 billion total parameters, 42 billion active at a time, natively multimodal, with a 1 million token context window. It is aimed at coding, long-running agent tasks, and multi-step reasoning.
The headline trick is where it runs. Meta optimized it for Apple's MLX framework and says it runs on two Mac Studios with the M5 Ultra chip. Each of those can give its GPU up to 512 GB of unified memory, so you are looking at roughly a terabyte of memory between the two boxes. Meta showed it paired with Muse Code, its coding terminal, fixing bugs and adding features to the Ice Cubes Mastodon app, and reviewing and cropping footage inside DaVinci Resolve, all on-device.
If you are not rocking two maxed-out Mac Studios, the weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. You can hit it by API through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. It ships under the Muse Model License, which Meta describes as having limited commercial terms but free for a lot of developers. Meta's page has a benchmarks section, but the published text did not give scores or name the models it beat, so I am not going to repeat a claim I cannot check.
What I want to know is who this is really for. "Runs locally" is doing a lot of work when local means two machines that cost more than my truck. Still, I like the direction. A frontier-ish model you can download, read the card on, and run without anybody's API meter ticking is good for builders and good for privacy. I just want to see the benchmark numbers before I believe the "most capable" line.
Source: Meta's Muse Spark Open announcement.
03 of 03
Anthropic gives METR an employee badge and gives up its veto
Anthropic announced Thursday that six researchers from METR, the nonprofit that tests frontier AI models, now have what it calls "employee-like access." That means Anthropic-issued hardware, office space, and access to internal tools roughly on par with Anthropic's own risk assessment teams. Anthropic says its staff are encouraged to talk to them like they would a colleague.
The part with teeth: METR will publish its own analysis of Anthropic's risk levels, incidents, and practices, and Anthropic cannot block it. Anthropic can only require redactions in predefined categories, and the examples it lists are customer data, attorney-client privileged material, and model architecture details. METR can then say publicly whether it thinks a redaction was justified.
Anthropic laid out three tiers of outside evaluators. Tier 1 is narrow investigations where Anthropic controls access and approves publication. Tier 2 gets more info and can publish without approval. Tier 3 is the employee-style setup, and METR is the only one there. SecureBio is reviewing the chem and bio weapons chapter of Anthropic's August 2026 Risk Report under more limited terms. Accenture and Faculty are working toward a broader partnership, and Hugging Face is in early talks. This follows Dario Amodei's September essay arguing for embedded evaluators, and a White House accord on independent review signed by OpenAI, Anthropic, Google, Meta, Nvidia, and xAI.
I will give credit where it is due. Giving up the veto is the whole ballgame here, because an audit the company can kill is just marketing. The test is what METR actually publishes and whether it ever stings. If the first report comes back all sunshine, folks will call it theater. If it names a real problem and Anthropic eats it, then we have something every other lab ought to copy.