OpenAI quietly changed what most people experience as AI. GPT-5.5 Instant is now the default model in ChatGPT — not the most powerful model in OpenAI's lineup, but apparently the one they trust most to put in front of hundreds of millions of users. The headline claim is reduced hallucination in sensitive domains: law, medicine, finance. Not improved creativity or longer context. Better truthfulness in the domains where being wrong has real consequences.

That is a deliberate and interesting bet. The name "Instant" signals speed, but the real pitch is precision in high-stakes situations. This mirrors years of enterprise feedback: customers do not primarily want an AI that writes better poetry — they want one that will not confidently cite cases that do not exist or fabricate drug interactions. Defaulting to a model that explicitly addresses this over one that maximises benchmark scores suggests OpenAI believes reliability is now the primary purchase criterion at scale.

The timing is layered. This week, Elon Musk and Sam Altman are in trial in Northern California over OpenAI's for-profit conversion and the terms of its impending IPO. An AI executive reportedly testified that Musk "was going to hit" him. A model release that reduces medical and legal hallucination is exactly the kind of product signal that makes a for-profit OpenAI look like a mature, enterprise-ready company rather than a research lab operating outside normal accountability structures. The court case and the product launch are separately motivated but together they tell the same story: OpenAI is professionalising rapidly and publicly.

Also this week: Freshworks cut 500 jobs — 11% of its workforce — while reporting $228 million in Q1 revenue. The company explicitly cited AI as the structural driver. This is the emerging pattern. AI does not reduce revenue; it reduces the headcount required to generate a given level of output. Every percentage point of hallucination reduction in a model like GPT-5.5 Instant makes that substitution safer for enterprises to act on at greater scale. The Freshworks number and the OpenAI release are not unrelated stories.

The less-noticed part of today is Anthropic's parallel work on sycophancy — building classifiers to measure whether Claude maintains positions under challenge and gives proportional praise rather than empty validation. Two of the three major frontier labs are now explicitly working on making their models less agreeable and more honest. That convergence matters. An AI ecosystem that reduces both hallucination and flattery is moving toward something qualitatively different from what we had eighteen months ago: not just more capable, but more genuinely trustworthy.