GPT-5.1 drops: Less thinking, more talking

OpenAI just dropped system card details for GPT‑5.1 Instant and GPT‑5.1 Thinking, and the headline isn’t raw horsepower — it’s how these models decide when to actually use their brains. GPT‑5.1 Instant now ships with what the company calls adaptive reasoning. Translation: the model figures out on its own whether a query deserves a quick chat response or a deeper think. GPT‑5.1 Thinking takes the same idea further, adjusting thinking time with more granularity per question.

The practical upshot for most users? You probably won’t touch a model selector. GPT‑5.1 Auto continues routing queries behind the scenes, picking the right engine without asking permission. OpenAI’s framing is that users “in most cases” won’t need to choose at all. That’s the real product story here — the model picker is becoming infrastructure, not a feature.

Safety-wise, the company is expanding what it checks before shipping. The addendum references new baseline evaluations covering mental health scenarios — specifically, situations where a user shows signs of isolated delusions, psychosis, or mania — plus a category they’re calling emotional reliance, which flags unhealthy attachment to ChatGPT. These join the existing battery described in the full GPT‑5 System Card. OpenAI says the comprehensive mitigations haven’t fundamentally changed; they’re just measuring more things now.

I’ll say what the system card doesn’t: this reads like a maturity play. The “adaptive reasoning” label is doing a lot of work to paper over what’s essentially a latency-versus-depth tradeoff dressed up as intelligence. Still, baking mental health and emotional dependency checks into pre-deployment review is a genuine expansion of scope. Whether those evaluations catch edge cases in the wild is the question no system card can answer.