Chatbots Are Becoming Products, Not Just Models: OpenAI's Legal Configuration, GPT-5.5's Six-Month Lifespan, and Oregon's Companion Chatbot Law

Three unrelated announcements landed this week, and none of them was a new frontier model. A law-firm product, a retirement notice, and a state statute — together they say something a benchmark chart never does: the hard part of building on chat models in 2026 isn't getting a smarter model, it's everything downstream of that — configuring it for a domain, keeping up as it gets swapped out from under you, and proving to a regulator that it knows when to stop talking.

1. A Legal Chatbot That's Explicitly Not a New Model

On September 17, OpenAI introduced Astra for Law, and the framing is the interesting part: it's GPT-6 Astra paired with a legal search index spanning U.S. case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, plus custom instructions for legal analysis, drafting, and issue-spotting. No new base model, just the existing one wrapped in retrieval, prompting, and governance built for one profession. On OpenAI's own evaluation, that combination passed 54.0% of questions on overall correctness at top reasoning effort, versus 38.7% for GPT-6 Astra using plain web search — a real jump, but one that came entirely from what sits around the model, not from the model itself. Firms get in through a Trusted Access program, 26 new plugins connect it to tools they already run like Relativity and Clio, and legal-tech vendors Harvey and Legora can build the same configuration into their own products via API. The takeaway for anyone shipping a chatbot into a specific field: the differentiation is shifting to the retrieval layer, the eval harness, and the integrations — the model underneath is table stakes.

2. Six Months Is Now a Model's Full Lifespan

GPT-5.5 launched on April 23, 2026. OpenAI confirmed this week it retires from ChatGPT, ChatGPT Work, and Codex on October 14 — under six months from ship to sunset in the consumer and business surfaces, with users pushed toward GPT-5.6 Sol or GPT-6 Astra. The API is the one exception: developers who call GPT-5.5 directly keep access, which is precisely the gap that matters. A team that built a product against "whatever ChatGPT currently is" just lost their model with five weeks' notice; a team that pinned a specific model ID in the API did not. That split is becoming the practical argument for calling models through an API with an explicit version rather than depending on a consumer surface's current default — not a best practice from a blog post, but a scheduling risk that already happened to whoever shipped on GPT-5.5 this spring.

3. Oregon Makes a Chatbot Stop Talking

Oregon's SB 1546, signed in March and taking effect January 1, 2027, joins California and Washington in regulating "AI companions" — systems designed to simulate a sustained, human-like platonic, intimate, or romantic relationship. What makes it sharper than California's SB 243 is one word: interruption. Where California requires a crisis-referral protocol, Oregon requires operators to detect suicidal ideation or self-harm language and actively interrupt the conversation to hand the user to the 988 Suicide & Crisis Lifeline or the Youthline, not just append a resource link and keep going. Operators also owe an annual public filing on how often that happened and how it shaped their intervention design, and the law carries a private right of action at $1,000 per violation — a plaintiff's attorney doesn't need to prove damages, just a missed interruption. For anyone building a bot with sustained, emotionally toned conversations, "detect and redirect" is no longer a nice-to-have moderation layer; in Oregon, it's the one feature a missed release can get you sued over.

The Common Thread

None of this week's news was about raw model capability. It was about what has to exist around a capable model before it's a shippable product: a retrieval and eval layer that turns a general model into a legal one, a versioning discipline that survives a vendor retiring your model on six weeks' notice, and a real-time safety intervention a regulator will now audit. The chat model is turning into a commodity input; the product is the configuration, the pinning strategy, and the compliance layer wrapped around it.

// A minimal fallback chain so a retired model doesn't take the product down with it
const MODEL_FALLBACK_CHAIN = ['gpt-5.5', 'gpt-5.6-sol', 'gpt-6-astra'] as const;

async function callWithFallback(prompt: string, chain = MODEL_FALLBACK_CHAIN): Promise<string> {
  for (const model of chain) {
    try {
      return await callModel(model, prompt);
    } catch (err) {
      if (!isModelRetiredError(err)) throw err; // only fall through on a deprecation, not a real failure
    }
  }
  throw new Error('All models in fallback chain are unavailable');
}

Suggested visuals: a simple before/after diagram showing GPT-6 Astra as the base model with Astra for Law's retrieval index, instructions, and plugins layered on top; a timeline bar comparing GPT-5.5's roughly six-month consumer-surface lifespan against older OpenAI model retirements to show the cycle shortening; and a side-by-side of California SB 243 versus Oregon SB 1546 highlighting the "referral" vs. "mandatory interruption" distinction for a compliance-focused reader.

— Maya