Tech · · Yunsuk Choi
1. Product context
On May 5, OpenAI replaced ChatGPT's default model with GPT-5.5 Instant. The headline change: hallucinations dropped 52.5% in sensitive domains — medicine, law, and finance — while speed stayed at 5.3 Instant levels.
*Photo by Jonathan Kemper on Unsplash*
2. What changed
Putting together OpenAI's launch post, TechCrunch, and Axios:
| Item | GPT-5.3 Instant | GPT-5.5 Instant |
|---|---|---|
| Released | Previous default | 2026-05-05 |
| Hallucinations (medicine, law, finance) | Baseline | −52.5% |
| Speed | Fast | Maintained |
| Personalization | Limited | Past chats, files, Gmail integration |
| Availability | ChatGPT default | ChatGPT default (instant swap) |
The core message: accuracy up, speed unchanged.
3. Why a 52.5% drop matters
"What matters here is that the hallucination reduction comes from evaluation benchmarks, not marketing math. In trust-critical domains like medicine, law, and finance, that lands harder."
— paraphrased from SiliconANGLE
OpenAI calls out three domains specifically:
- Medicine — drug information and diagnostic support (not a substitute for medical advice)
- Law — statute interpretation and case citation
- Finance — general tax and investment information (not a substitute for licensed advice)
Fewer hallucinations does not mean it replaces a professional. But for research support, summarization, and first-pass document work, the reliability bar has moved up a notch.
4. Stronger at multi-step work
The other notable upgrade is GPT-5.5's ability to carry complex, multi-step tasks through to the end:
- Writing and debugging code
- Online research with structured output
- Data analysis through chart generation
- Drafting documents and spreadsheets
- Operating software (app and API calls)
We're moving from "one prompt, one task" to "throw it the goal and let it sequence the steps." Agentic usability is the headline.
5. Personalization — Gmail in the loop
"ChatGPT can now reference past chats, uploaded files, and a connected Gmail account when responding. The rollout starts with Plus and Pro, then expands to Free and Business in stages."
— paraphrased from OpenAI's official post
- Plus and Pro users: live as of May 5
- Free, Go, Business, Enterprise: phased rollout over the coming weeks
- Mobile: coming soon
The real point of better personalization is reusable context. You stop re-explaining yourself every session.
6. How it stacks against Claude and Gemini
A cluster of model launches landed in the same window:
| Model | Strength | Released |
|---|---|---|
| Claude Opus 4.7 | Coding, long-horizon work, agents | May 18 |
| Gemini 3.5 Flash | Speed (12×), half the cost | May 20 (Google I/O) |
| GPT-5.5 Instant | Lower hallucinations, multi-step | May 5 |
OpenAI is leading on accuracy. Speed belongs to Gemini; coding and long-horizon tasks belong to Claude. A year ago, one model could carry everything. Now you pick a model per task.
7. What ChatGPT users should do
- Better on sensitive questions — but still no substitute for medical, legal, or financial professionals
- Lean on personalization — store the contexts you reuse most
- Delegate multi-step work — try compound asks like "tidy these emails, summarize them, then add the meetings to my calendar"
- Free users: the upgrade reaches you too in phases
- API pricing: check OpenAI's separate API pricing notice for GPT-5.5
*Photo by Aditya Parikh on Unsplash*
8. TL;DR
- May 5, 2026: GPT-5.5 Instant becomes ChatGPT's default
- Hallucinations cut 52.5% in medicine, law, and finance
- Speed stays at 5.3 Instant levels
- Personalization (past chats, files, Gmail) rolling out in stages
- Stronger at multi-step tasks
- Distinct strengths now versus Claude and Gemini
The biggest shift is that GPT's old weak spot — accuracy — is closing fast. But for medicine, law, and finance, the expert consult still comes first.
9. Related tech notes
For a related thread, see the IT category or under #AI and #ChatGPT. Also see Google I/O 2026 Gemini 3.5 recap.
10. Sources
Sources: OpenAI — GPT-5.5 intro, OpenAI — Instant details, TechCrunch, Axios, SiliconANGLE
Tags: #OpenAI #ChatGPT #GPT5 #AI