OpenAI Updates ChatGPT with GPT-5.5 Instant, Fewer Hallucinations
OpenAI has switched ChatGPT's default model to GPT-5.5 Instant. This change cuts down on hallucinations and improves response accuracy. A fresh tool named memory sources lets users view the stored details that influence replies.
GPT-5.5 Instant takes over from GPT-5.3 Instant. Developers can access it via the API under the name "chat-latest." Company tests showed it makes 52.5 percent fewer false statements than the prior version on sensitive prompts related to medicine, law, and finance. For user-marked chats with past errors, wrong claims fell by 37.3 percent, according to OpenAI.
Improvements in Reasoning and Error Detection
OpenAI shared an example with algebra. Someone uploaded a picture of a handwritten equation that had a wrong calculation. GPT-5.3 Instant first matched the user's answer. It saw x=3 failed but said no real solution existed. GPT-5.5 Instant started the same way. Then it spotted the mistake in rearranging the equation and fixed the quadratic properly.
OpenAI, founded in 2015 as a nonprofit focused on safe artificial general intelligence, launched ChatGPT in November 2022. That free tool quickly gained over 100 million users. The GPT series has advanced through versions like GPT-4 and GPT-5, powering applications in text, code, and images.
Benchmark Performance Gains
Tests confirm the progress. On AIME 2025, a math competition exam, scores rose from 65.4 percent to 81.2 percent. GPQA, testing PhD-level science questions, improved from 78.5 percent to 85.6 percent. CharXiv, for reasoning on scientific charts, went up from 75.0 percent to 81.6 percent.
MMMU-Pro, evaluating expert questions with text and images, climbed from 69.2 percent to 76.0 percent. OmniDocBench error rate for pulling data from complex documents dropped from 14.6 percent to 12.5 percent.
| Benchmark | Description | Metric | GPT-5.3 Instant | GPT-5.5 Instant | |, , , , , -|, , , , , , -|, , , , |, , , , , , , , -|, , , , , , , , -| | CharXiv-reasoning | Scientific Chart Reasoning | Accuracy | 75.0% | 81.6% | | MMMU-Pro | Expert Multimodal Reasoning | Accuracy | 69.2% | 76.0% | | OmniDocBench | Document Parsing | Average error rate (lower = better) | 14.6% | 12.5% | | GPQA | PhD-Level Science | Accuracy | 78.5% | 85.6% | | AIME 2025 | Competition Math | Accuracy | 65.4% | 81.2% |
Shorter Responses and Smarter Use of Context
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
OpenAI trimmed extra words. Replies stay brief yet complete. The model skips pointless questions, extra emojis, and excess styling. "It can deliver the same information, often with more utility than previous models, while reducing the verbosity and overformatting that can make responses too long," OpenAI states.
GPT-5.5 Instant handles context from old chats, files, and linked Gmail better when enabled. It decides wisely if personalization fits. It also scans past talks quicker.
Memory Sources Across ChatGPT
Memory sources now work on all ChatGPT models. Users see exactly which saved notes, past chats, or files shaped a reply. They can mark items as useful or not, change them, or erase them.
Not every influence appears, OpenAI notes. Only select searched chats show up. Plans call for fuller views later. Shared chats skip these sources. Temporary chats ignore and avoid memory.
OpenAI brought out GPT-5.5 Thinking as the premium option recently. GPT-5.5 Instant fits daily use. The Thinking model beats on cybersecurity like Claude Mythos and swaps out Codex for coding.
Rollout to Users
GPT-5.5 Instant reaches all ChatGPT users now. Paid accounts keep GPT-5.3 Instant in settings for three months until retirement.
Deeper personalization with chats, files, and Gmail starts for Plus and Pro on web. Mobile follows soon. Free, Go, Business, and Enterprise get it in weeks. Memory sources hit consumer plans on web first, then mobile. Some features skip certain regions.