SpaceXAI just released Grok 4.5. It may have broken California’s AI law in the process.

On Wednesday, SpaceXAI (formerly xAI) launched Grok 4.5, its most capable model to date. But what didn't ship was either a transparency report or a model card for the new model. 

Under a California law that’s been in effect since the beginning of the year, that may be a problem.

California’s SB 53, passed in September of last year, requires companies like SpaceXAI to publish transparency reports with every major model release. These reports generally take the form of system cards. In addition to intended uses and restrictions, these reports must include summaries of risk evaluations conducted on the model, the results of said evaluations, the extent to which third-party evaluators were involved, and any other steps taken to comply with the developer’s frontier AI framework. 

The system cards released by leading AI companies, including OpenAI and Anthropic, often total dozens or hundreds of pages. Meta’s newest model, released today, features a 112-page system card of this sort. SpaceXAI, however, has long been an outlier. Of the nearly 20 models that SpaceXAI has released, only a handful have been accompanied by anything resembling a model card, and these have sometimes been quite sparse.

The requirement to disclose a model card or equivalent is now active and enforceable under California law. Was SpaceXAI required to release one for Grok 4.5? It sure looks like it. 

There are a few qualifying requirements to determine whether a model is covered under SB 53. The first is that the company is a large frontier developer, meaning it earns gross revenues in excess of $500 million a year. According to the recent SpaceX IPO filing, SpaceXAI recorded $3.2 billion of revenue in 2025, satisfying this first requirement.

The second requirement is that the model be a frontier model, which is to say that it was trained with a quantity of computing power exceeding 10^26 floating-point operations (FLOP). Without insider details about the training of Grok 4.5, it’s hard to say definitively whether it meets this requirement, but most signs point toward yes.

Grok 3, released in February 2025, is the first model that Epoch AI estimated to exceed the 10^26 FLOP threshold. Grok 4, released later that year, was estimated by Epoch to have exceeded the threshold by a factor of five. While recent trends in model training have sometimes entailed newer models using less total compute than their predecessors, it seems unlikely that Grok 4.5 (built on top of a brand-new, 1.5 trillion-parameter pretrain and trained across “tens of thousands of Nvidia GB300 GPUs”) experienced this downgrade to an extent that would have left the final post-trained model below the 10^26 FLOP threshold.

The last argument that SpaceXAI could make to contend that Grok 4.5 is not a covered model under SB 53 would be that the 4.5 model is based on earlier models in the Grok 4 line and therefore is not new, or in the law’s words, “substantially modified.” But this seems unlikely, and not just because of the new tier of capability improvements reached by 4.5. Elon Musk himself has admitted that the base model used for Grok 4.5 is different from the model used for Grok 4.3, the previous entry in the Grok 4 generation. In other words, despite the naming similarities with past SpaceXAI releases, Grok 4.5 represents a genuinely new model built from the ground up.

The safety page on SpaceXAI’s website, where the company details its safety practices and publishes its model cards (but which is also 100% AI-generated), currently makes the broad claim that SpaceXAI publishes “detailed model cards for each of our frontier models.” But, as mentioned earlier, SpaceXAI has not released system cards for the majority of the models it has released to date. In fact, in recent testimony, Elon Musk claimed under oath that he was “not sure what a safety card is.”

Will a system card eventually be released for Grok 4.5? It could happen. Grok 4’s eight-page system card was released six weeks after the release of the model itself. But SB 53’s requirement is that the system card be published “before, or concurrently with” the release of the model, meaning that even a weeks-delayed system card may still represent a violation of California’s new law.

The Midas Project has previously raised questions about the quality of SpaceXAI’s safety practices. We’ve also raised questions about potential early SB 53 compliance missteps from OpenAI and Google. But those missteps were small compared to this: the apparent absence of any serious safety effort for a brand-new model marketed by the company as approaching Fable-level coding performance.

To put a finer point on it, the Trump administration essentially demanded the removal of Anthropic’s Fable model from the market due to a jailbreak, even though the model was released with hundreds of pages of documented safety testing determining it to be fit for public access. Grok 4.5, meanwhile, is marketed as outperforming Fable on at least one coding benchmark, was released with no documented safety testing at all, and, only hours after release, was jailbroken, generating a Remote Access Trojan script alongside other harmful outputs.

I’m not holding my breath for the Trump administration to remove Grok from the market (for many reasons, including that I don’t believe Grok 4.5’s performance is actually competitive with Fable). 

However, it does seem like a prime opportunity for California to show that its new AI safety law — a minimally burdensome statute asking only for transparency into the nature and extent of the safety testing conducted by the company — actually has teeth.