New York Restricts Companion Chatbot Features for Minors, ChatGPT Dot Moves Into Text Messages, and GPT-5.5 Gets Retired

Three stories from the last few days: a state moving from lawsuits to statute on companion chatbots, a proactive assistant leaving the app for the text-message thread, and a reminder that the model you built on has a shelf life. Details come from news reports and search summaries I could not open in full, so treat specifics as the reporters' own.

1. New York Regulates Chatbot Features, Not Just Chatbots

Reports dated October 28 say Governor Kathy Hochul signed two bills aimed at technology in children's bedrooms: one on AI chatbots, one on internet-connected baby monitors. According to those reports, the chatbot law bars operators from offering a defined set of companion-style features to anyone under 18, or to any user whose age the operator hasn't confirmed, with civil penalties reported at up to $25,000 per violation and an effective date of January 1, 2027. Earlier coverage of the bill describes the targeted behaviors as simulated companionship, sexually explicit interaction, self-harm promotion, secrecy from caring adults, and sycophancy used to keep young users engaged. It does not ban minors from using chatbots outright. I could not confirm the exact statutory text or penalty details, so read the bill before relying on any of this.

The notable design choice is the phrase "or whose age isn't confirmed." That makes unknown age the same as minor, which turns age assurance into a prerequisite for shipping persona, memory or relationship features at all. It lands alongside the state attorney-general suits against OpenAI and a stalled federal bill, so state-by-state feature gating is the realistic compliance model for now.

A practical pattern is to gate by capability, not by product:

type AgeStatus = 'adult_verified' | 'minor' | 'unknown';

function featuresFor(age: AgeStatus) {
  const adult = age === 'adult_verified';
  return {
    persona: adult,
    longTermMemory: adult,
    proactiveNudges: adult,
    coreAnswers: true, // always on
  };
}

Default to the restricted set, and make the verified path an explicit upgrade you can audit.

2. Proactive Agents Move Into the Messaging Thread

9to5Mac reported on October 28 that OpenAI is extending ChatGPT dot, described as its proactive agent, to text messaging: users can set tasks and receive updates over iMessage or RCS. This rests on a single report plus a ChatGPT social post it embeds; I found no OpenAI press release. Separately, release notes logged on October 27 show ChatGPT connecting to iCloud Mail and Calendar for searching email and updating events.

For builders the signal is channel, not novelty. Proactive agents that push updates need delivery you don't control: carrier rules, opt-in and opt-out handling, rate limits and quiet hours. They also raise the stakes on the New York-style rules above, because an agent that texts first is exactly the kind of engagement feature regulators single out. If you add outbound messaging, store consent per channel, honor STOP everywhere, and log every proactive send with the reason that triggered it.

3. Your Model Has an Expiry Date

OpenAI's model notes say GPT-5.5 is being retired from ChatGPT, ChatGPT Work and Codex, with API access unaffected. Releasebot also logged GPT-6.1 Terra arriving in the API, Work and Codex on October 26, but not in standard chat. Whatever the exact rollout, the pattern is that ChatGPT-side model lines turn over in weeks, and surfaces within one vendor get different models.

Pin model IDs in config, not code, and keep a small regression set of real conversations you replay before changing the default. Compare tool-call behavior and refusal rates along with answer quality; those are the things that quietly shift between versions.

Conclusion

Regulation is targeting features, assistants are moving into channels with their own rules, and models rotate underneath you. The common defense is boring engineering: capability flags tied to verified age, per-channel consent logs, and a replayable eval suite. Build those once and each of these headlines becomes a config change instead of a rewrite.

Suggested visuals: a feature-by-age-status matrix for the gating example, and a timeline of model retirements against your own pinned versions.

— Maya