OpenAI makes GPT-5.5 Instant the default ChatGPT model, cuts hallucinations

OpenAI has replaced ChatGPT’s default model with GPT-5.5 Instant, available via the API as "chat-latest." Internal tests show 52.5% fewer hallucinated claims on high‑risk prompts and a 37.3% reduction in inaccurate claims on previously flagged conversations.

OpenAI has replaced ChatGPT’s default production model with GPT-5.5 Instant and is offering the model through its API under the label "chat-latest." The company says the swap is live for ChatGPT users immediately and that GPT-5.5 Instant succeeds GPT-5.3 Instant as the standard production model.

In internal evaluations, OpenAI reports that GPT-5.5 Instant produced 52.5% fewer hallucinated claims than the prior release on high‑risk prompts covering medicine, law and finance, and it cut inaccurate claims on previously flagged user conversations by 37.3%. OpenAI illustrates the change with a handwritten algebra example: where GPT-5.3 Instant initially agreed with a user’s incorrect rearrangement and later concluded no real solution existed, GPT-5.5 Instant caught the rearrangement error and solved the corrected quadratic.

Benchmarks show broad gains across math, science and multimodal reasoning tasks. Accuracy on AIME 2025 rose from 65.4% to 81.2%; GPQA (PhD‑level science) increased from 78.5% to 85.6%; CharXiv (chart reasoning) climbed from 75.0% to 81.6%; MMMU — Pro moved from 69.2% to 76.0%; and OmniDocBench’s average error rate fell from 14.6% to 12.5%, according to OpenAI.

The update also introduces a "memory sources" feature that surfaces which stored context — saved notes, past chats, uploaded files or reminders — contributed to a particular reply. Users can flag entries as relevant or irrelevant, edit them or delete them. OpenAI cautions the view is not exhaustive today (only some searched chats will appear) and that memory sources are not included when a chat is shared; the company says it plans to expand this transparency over time.

OpenAI says GPT-5.5 Instant produces shorter, less overformatted replies with fewer emojis and unnecessary clarifying questions, and it improves judgment about when personalization (drawing on past chats, files or connected Gmail) is helpful. That personalization and access to memory sources will roll out first to Plus and Pro web subscribers, with broader availability planned in the following weeks.

For builders and integrators, the immediate operational changes to test against are the default model swap and the API label "chat-latest." The reported accuracy and benchmark gains reduce some exposure to high‑risk hallucinations, while memory sources add a layer of auditability but remain incomplete and are excluded from shared chats today — a detail that matters for apps that log or hand off conversations. Developers should validate changes in verbosity, personalization and document parsing against their own workloads.