OpenAI swaps the ChatGPT default model — GPT-5.5 Instant replaces 5.3
On 5 May 2026, OpenAI rolled out GPT-5.5 Instant as the new default model for ChatGPT. The model replaces GPT-5.3 Instant in the free web and app versions and is also available via the API as chat-latest. The most visible difference in practical testing: significantly fewer fabricated facts in sensitive areas like law, medicine and finance — and notably fewer emojis in the response style.
What changed in concrete terms
- Default model swap: GPT-5.5 Instant replaces GPT-5.3 Instant for all ChatGPT users, including the free tier. The API endpoint is still
chat-latest, but it now points to 5.5 Instant. - -52.5% hallucinations in OpenAI’s own evaluations versus 5.3 Instant — tested against high-stakes prompts on medicine, law and finance.
- Fewer emojis, tighter responses: OpenAI itself calls it “less gratuitous emoji use”; answers stay closer to the point.
- Memory sources are now visible: ChatGPT now displays — across all models — where a personalized answer was generated from. Sources can be deleted or corrected.
- Personalized search referencing past chats, files and Gmail is initially Plus/Pro on web only; mobile, Free, Go Business and Enterprise follow in the coming weeks.
What was the case before
GPT-5.3 Instant had been ChatGPT’s fast default model since early 2026. Its job was clear: low latency for everyday queries, while heavier reasoning models lived behind the picker. Known criticisms: noticeable hallucinations on factual queries in law, medicine and finance — and a response style that drowned users in emojis without anyone asking for them.
On the API side, GPT-5.3 Instant was served via the chat-latest endpoint. Anyone using it automatically got the current default model without having to deal with model IDs.
What’s the case now
1. GPT-5.5 Instant is the new default — including the free tier. Unlike many Pro-only releases, OpenAI is rolling the new model out to every ChatGPT tier at once. Anyone who doesn’t actively change the picker has been getting GPT-5.5 Instant since 5 May 2026. The API endpoint name does not change: chat-latest now points to 5.5 Instant. To pin a specific model, set the explicit model ID.
2. Hallucination reduction in sensitive domains. The headline number — 52.5% fewer hallucinated claims versus 5.3 Instant — comes from internal OpenAI evaluations and specifically targets prompts on medicine, law and finance. That’s the domain where hallucinations are most expensive: wrong dosages, fabricated statutes, invented tax rates. The number is not externally reproducible because OpenAI does not release the eval set — but it lines up with hands-on tests by tech reporters (TechCrunch, 9to5Mac), who saw fewer fabricated sources in comparable prompts.
3. Memory sources are visible — and editable. ChatGPT’s memory used to be a black box: it would remember things without showing where they came from. With the 5.5 Instant update, ChatGPT now shows — across all models — which source (chat, file, Gmail) a personalized answer was drawn from. Stale sources can be deleted, incorrect ones can be corrected. This is the more important change, even though the press release foregrounds the hallucination number.
4. Anti-emoji style correction. OpenAI puts it cautiously (“less gratuitous emoji use”), but the reporting is clearer: answers come decorated with fewer emojis and stay tighter and more matter-of-fact. This is not pure cosmetics — many professional users paste ChatGPT output directly into emails, documentation or tickets, where emojis are noise.
Why it matters
Default-model swaps are the highest-leverage moves, because 90% of users never touch the picker. Anyone using ChatGPT as an information-retrieval tool — quick research on legal questions, medical symptoms, tax situations — has been getting measurably more reliable answers since 5 May, without having to change anything.
What’s more interesting is what the change does not do: GPT-5.5 Instant remains the fast default — no reasoning latency, no expensive tool-use loops. Anyone who wants full GPT-5.5 or GPT-5.5 Pro for complex tasks still has to switch manually. The default swap raises the floor without touching the ceiling.
For GEO/AI-search observers: hallucination reduction in high-stakes domains also means ChatGPT cites more accurately in exactly those areas. Sites trying to establish themselves as sources for medical, legal or financial content now face slightly more pressure to be cleanly structured — because the model is more critical of bad citations.
What you can do now
If you use ChatGPT for research: No action needed — GPT-5.5 Instant is the new default. But spot-check critical answers, especially if you’ve been relying on 5.3 Instant outputs. The swap can shift answers noticeably in sensitive domains.
If you use ChatGPT via the API: If your app points to chat-latest, you’ve been on GPT-5.5 Instant since 5 May. Pin an explicit model ID if you need reproducibility — e.g. for eval pipelines or audit logs.
If you care about memory sources: Open your ChatGPT profile, review the memory entries, and delete anything outdated. The new visibility only helps if you actually use it.
Entdecke mehr
Three Days with Claude Fable 5 — and Why Opus Stays My Daily Driver
Fable 5 fixed a bug in one shot that Opus failed at twice. I'm still not switching. My honest impressions from the trenches after three days.
GlossarGPT-4o
GPT-4o (May 2024) was OpenAI's first natively multimodal model — text, audio and image in a single model, with a real-time voice mode and substantially lower cost than GPT-4 Turbo.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.