Eight States Sue OpenAI Over Child Safety, GPT-6 Splits Into Sol and Luna by Plan, and Google's Work Agent Routes Each Job to Gemini or Claude

Three stories this week share a theme: the model behind a chatbot is no longer one fixed thing, and the people responsible for it are being asked to account for that. Attorneys general want to know what a product told families about safety, OpenAI now serves different GPT-6 models by plan, and Google is previewing an agent that chooses its own model per task.

1. Eight States Sued OpenAI in a Single Day

On October 14, attorneys general from Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana each filed suit in state court. Per CNBC, NBC News and local coverage, the complaints allege ChatGPT harmed children and that OpenAI misled users about its safety, under state consumer-protection law. The states are asking courts to order changes to how ChatGPT works for minors and to make OpenAI give up profits earned from young users. Florida's attorney general filed a similar suit in June, and California courts already hold seven suits alleging wrongful death and negligence. OpenAI says it will keep building safeguards such as parental controls and break reminders.

The legal theory is worth noticing. These are consumer-protection claims about what the product said about itself, not claims that a model output was unlawful. For anyone shipping a chatbot, that makes your safety copy, age-gating flow and marketing claims part of the attack surface. If your landing page says "safe for teens", your logs need to show what that means in practice.

2. GPT-6 Arrives as Two Models, Plus Answers That Are Also Apps

ChatGPT's GPT-6 rollout started October 7 with paid tiers on GPT-6 Sol, and Free and Go users on GPT-6 Luna the next day. It came with "Intelligent UI", where an answer can include adjustable charts, forms, and small interactive tools such as calculators or bill splitters, generated inside the conversation.

For developers, two things matter. First, "which model answered?" now depends on the user's plan, so evaluation has to be per tier. A prompt that works on the flagship may fail on the cheaper model your free users get. Second, interactive answers move the chat surface from text toward a rendering target. If you plan something similar, treat model-generated UI as untrusted input: constrain it to a fixed component schema instead of letting the model emit arbitrary markup.

3. Google's Work Agent Picks Its Own Model

At Gemini at Work 2026 on October 8, Google Cloud previewed a universal work agent with one prompt box that runs jobs on either Gemini or Anthropic's Claude models, choosing whichever suits the task. It is in private preview with no pricing, so treat the details as early. A model-agnostic agent from a company with its own frontier model is a notable signal: routing is becoming the product, and the model underneath is a replaceable part.

A minimal version of that idea is small enough to sketch:

type Route = { model: string; reason: string };

function pickModel(task: { needsTools: boolean; tokens: number }): Route {
  if (task.needsTools) return { model: 'agent-tier', reason: 'tool use' };
  if (task.tokens < 2000) return { model: 'small-fast', reason: 'short, cheap' };
  return { model: 'flagship', reason: 'long or complex' };
}

Whatever you route on, log the choice with the response. When a user complains, "which model said that" should be a query, not a guess.

What Connects the Three

A regulator asks what you promised, a vendor ships different models to different users, and an agent chooses a model on its own. In each case the unit of accountability is the whole system, not a model name. Suggested visuals for this post: a table of GPT-6 tiers by plan, and a diagram of a routing layer with the audit log branching off it.

Conclusion

Record which model answered, for which user, under which safety claims. This week's news suggests all three will be asked about.

— Maya