Four nonprofit organisations brought the case: Public Citizen, The Heat Initiative, the Irreplaceable Education Fund and the Alliance for Secure AI. The filing was announced in a Public Citizen press release on 12 November 2026 and has been reported by Just Security and unite.ai.
What the injunction would cover
The plaintiffs are not asking for a fine or a labelling rule. They want a court order that would pause the training, development and deployment of frontier models while the case is heard.
As described, that request covers work on models not yet released as well as the services already on the market. The legal basis is California's Unfair Competition Law. The argument, in short: building highly autonomous systems without proven safety guarantees amounts to a public nuisance and an unfair business practice.
Four organisations against five companies
The defendants are OpenAI, Anthropic, Google's parent Alphabet, Meta and SpaceXSI. OpenAI, Anthropic, Google and Meta all sell AI services to European customers. SpaceXSI is named alongside them.
The count is unusual for this kind of case. Earlier legal pressure in the United States went after single labs, mostly OpenAI. This complaint treats the frontier labs as one group facing one shared claim.
The protests that came first
The lawsuit follows a summer of street action. Marches organised by Stop the AI Race drew an estimated 200 to 400 people outside the OpenAI, Anthropic and Google DeepMind offices in San Francisco, Decrypt reported. Kaufmyn's jailing in August produced the campaign's first arrest.
Four of the five defendants shipped new models
Between 4 and 12 November, four of the five companies named in the complaint released a new model. Exact per-token prices change frequently and are not reproduced here; the table records the releases and their positioning. Meta describes Muse Ember as its smallest, most efficient model built for high-volume agent work, so it is a lightweight high-volume model, and only three of the four releases — GPT-6 Cyber, Claude Fable 5.2 and Gemini 4 Hydrogen — are frontier-class.
| Model | Vendor | Release | Notes |
|---|---|---|---|
| GPT-6 Cyber | OpenAI | 4–5 Nov | Frontier-class release |
| Claude Fable 5.2 | Anthropic | 9 Nov | Frontier-class release |
| Gemini 4 Hydrogen | Google DeepMind | 12 Nov | Frontier-class release |
| Muse Ember | Meta | 8–9 Nov | Lightweight high-volume model |
| Mistral Large 4 | Mistral AI (not a defendant) | 5 Nov | European comparison |
Some vendors also advertise fast modes, free tiers and cached-input discounts; exact rates vary by model, volume and tier and should be checked on each vendor's current pricing page. European business customers are often billed in dollars for these APIs and settle VAT through the reverse-charge mechanism.
Europe's lever is a regulator, not a judge
The EU has no equivalent of this private lawsuit. What it has is enforcement. Since 2 August 2026, the European AI Office and national market-surveillance authorities have been applying Article 50 of the AI Act, the transparency provision. Providers must mark synthetic content with machine-readable tags, disclose deepfakes, and publish standardised summaries of their training data. Fines for Article 50 breaches fall under Article 99(4) and reach up to €15 million or 3% of global annual turnover, whichever is higher. The €35 million / 7% tier applies to prohibited practices under Article 5; the 7% ceiling belongs to those prohibited-practice breaches.
The two instruments work in opposite directions in time. A regulator can fine a company after a breach is found. The California plaintiffs want a judge to stop the work before it happens. Neither has halted a release so far, and all four AI companies named in the complaint place their models on the EU market, so the Article 50 duties apply to them directly.
What a European user would notice
Nothing yet. A complaint is not a ruling. No injunction has been issued, and the case has to reach a hearing before any order takes effect. The change this week is procedural: the demand for a freeze now sits on a court docket instead of on a protest permit.
For a teacher, a designer or a small firm paying for one of these APIs, the practical question is narrower. If an order were ever granted, it would bind the companies, not their customers, and the products in the table above are sold globally. That is the situation as filed. There is no schedule attached to it.
The filing itself runs to 63 pages. Four organisations signed it and five companies are named. In the two weeks before it landed, four of those five defendants released a new model. Meta's Muse Ember is a lightweight high-volume model; GPT-6 Cyber, Claude Fable 5.2 and Gemini 4 Hydrogen are the frontier-class releases.
Has a court granted the freeze yet?
No. The complaint has been filed in San Francisco Superior Court, but no injunction has been issued and no ruling has been reported. Until a judge acts, the five companies can keep training and deploying models.
Can the EU AI Act order a model pulled from the market?
The AI Act works through national market-surveillance authorities rather than private lawsuits. Those authorities can require a provider to correct a compliance failure and, in serious cases, restrict or withdraw a system. The route is regulatory. For Article 50 breaches, the fine ceiling is €15 million or 3% of global annual turnover; the 7% ceiling belongs to prohibited-practice breaches under Article 5.
Will API prices in Europe change because of the lawsuit?
Nothing in the complaint addresses pricing, and the filing is about training and deployment, not about what a token costs. Exact API prices should be checked with each vendor.
