GPT-5.5 Instant dropped May 5, 2026. 500 million ChatGPT users woke up with it by default. OpenAI’s own numbers claim 52.5% fewer hallucinations on high-stakes prompts. Medicine, law, finance. That number sounds like the problem is solved. It isn’t.
Hallucinations have been blocking enterprise AI adoption for two years.
Every team I know tried slapping AI into a customer-facing workflow and hit the same wall: the model sounded sure while outputting garbage. A lawyer drafting filings can’t ship confident nonsense. A doctor doing triage can’t either. Until this week, you just accepted the risk and kept a human babysitting every output.
This model also scored 81.2 on AIME 2025 math, up from 65.4.
MMMU-Pro multimodal reasoning hit 76.0 from 69.2. CharXiv scientific charts went 75.0 to 81.6. No latency penalty, supposedly. Those are real jumps — not marketing fluff.
The Feature That Nobody’s Talking About Is the One That Matters
Memory Sources. That’s what it’s called.
ChatGPT now shows you which past conversations, saved memories, or uploaded files shaped a given response. You can delete, correct, or flag anything. It’s rolling out to all consumer plans on web starting today. Enterprise and EDU get it later.
If you’re running AI workflows for clients.
Especially in healthcare, legal, or finance. This is the feature that changes your liability exposure. Your clients can finally see what the model “knows” about them. More importantly, they can see what it shouldn’t. That’s not a convenience feature. It’s an audit trail. “AI-assisted process” and “AI-generated liability” look identical to regulators. Memory Sources is what separates them.
Side note: OpenAI’s settings UI for this feature is buried three menus deep. You won’t find it by accident.
That tells you something about how eagerly they wanted users to manage this.
The $50 Billion Number Nobody Wants to Touch
Greg Brockman testified the same day.
Under oath. OpenAI will spend $50 billion on compute in 2026. Fifty billion. From $30 million in 2017.
That’s not a typo. That’s a 1,667x increase in nine years.
The enterprise handing you a “smarter” model told a courtroom it needs $600 billion by 2030 to cover its obligations.
Meanwhile, GPT-5.5 Instant is free to ChatGPT users and API pricing hasn’t budged.
OpenAI’s revenue is estimated $3-5 billion annually.
They’re burning 10 to 15 times intake. That’s not a business. That’s a compute subsidy program wrapped in a nonprofit governance structure that keeps getting re-litigated. That changes when the board changes, when investors get nervous, or when Washington decides AGI liability is a federal problem.
Copilot already killed flat-rate billing. June 1. Your $10/month buys $10 in tokens now. That’s not coincidence. Subsidize to lock you in. Tighten the meter once you’re dependent. I’ve seen this pattern with every cloud provider, every SaaS tool, every “introductory pricing” that converts to “standard pricing” six months later.
What You Do With This Information
Test the hallucination reduction against your actual workflows.
Don’t trust OpenAI’s internal numbers — verify them on your prompts, your edge cases, your domain. If you’re in medical or legal AI, the 52.5% reduction might matter enough to change your human-in-the-loop requirements. Or it might not. Test it.
Audit Memory Sources with your team this week. If you’re on a paid consumer plan, it’s already in your settings. Walk through it. Find the entries that shouldn’t be there. Delete them. Document what you found. If a client asks what the model knows about them, you need an answer that isn’t “I didn’t check.”
Build a dependency map.
What happens if OpenAI triples API prices next quarter? Do you have a fallback model that handles the same tasks? If the answer is no, that’s your next sprint item. Not next quarter. This quarter.
The hallucination number is probably real enough to matter. Memory Sources is genuinely useful. But the $50 billion number is the one that should keep you up at night if you’re running a business on someone else’s inference layer. OpenAI isn’t a utility. It’s a enterprise burning capital at a scale that requires either AGI economics to materialize or a structural bailout that makes the current investors whole. Neither outcome guarantees your API prices stay where they are.
Run multi-model. Audit your costs.
Check your Memory Sources settings today.
