OpenAI has begun rolling out GPT-5.5 Instant as the new default model inside ChatGPT, replacing GPT-5.3 Instant for users on free, Plus, Pro and Enterprise plans. The company is anchoring the launch to a single accuracy claim aimed at the regulated work that has historically been ChatGPT's weakest territory.
The headline number on this, that OpenAI keeps telling everyone, is that there is 52.5% fewer hallucinations, Jaeden Schafer said on the AI Chat Daily podcast, citing a write-up from Maxwell Zenith. OpenAI says that reduction is concentrated on what it calls high-stakes prompts — questions about medicine, law and finance, where confident wrong answers carry real downside for users and for the company's enterprise pitch.
The company is also reporting a 37% drop in inaccurate claims on the kind of tough conversations users had previously flagged, suggesting the model was tuned in part on its own logged failure cases. Standard benchmark scores moved with it: AIME math climbed from 65 to 81, GPQA from 78 to 85, and Charvi IV from 75 to 81.
Key facts
- 01OpenAI is rolling out GPT-5.5 Instant as the new default ChatGPT model across free, Plus, Pro and Enterprise tiers, replacing GPT-5.3 Instant.
- 02OpenAI claims 52.5% fewer hallucinations on high-stakes prompts in medicine, law and finance, plus 37% fewer inaccurate claims on tough conversations users had flagged.
- 03Benchmark scores jumped from 65 to 81 on AIME math, 78 to 85 on GPQA, and 75 to 81 on Charvi IV.
- 04A new memory sources panel lets users see which past chats, saved memories or Gmail data informed an answer, and delete or correct that context.
Schafer flagged the targeting as more interesting than the headline lift. "The thing that I do think is important is specifically at getting better at law, finance, and medicine. Those are some really critical areas that you can't have it messing anything up on," he said. Those verticals are also where OpenAI faces the heaviest competition for enterprise contracts and the highest liability exposure when models invent citations or misread statutes.
Alongside the model, OpenAI is shipping a memory sources panel that surfaces which pieces of stored context — past chats, saved memories, connected Gmail data — were used to personalize a given response. Users can delete or correct any of those entries directly from the panel, a notable shift for a feature that until now has largely operated as a black box.
Schafer said the transparency addresses a real failure mode in shared-account usage. He described situations where a friend or family member borrowed ChatGPT, asked something specific, and seeded the assistant with assumptions that bled into later sessions. "I let my friend ask a question to ChatGPT or ask a question to ChatGPT about his car. And now it's, you know, thinks this is my car forever," he said, calling the new ability to prune that context useful.
OpenAI is pairing the model release with renewed marketing around Codex, its coding-focused stack, where Anthropic's Claude has built a clear lead among developers. Schafer pointed to a post from Peter Gostef on X claiming a complex prompt run in Codex with GPT-5.5 was nailed in one shot, while noting he could not verify whether the post was sponsored.
Even with that caveat, Schafer said independent feedback on Codex has been trending positive, framing the launch as a coordinated push to claw back coding workloads. The combination of fewer hallucinations on regulated topics, visible memory controls and a sharpened coding model lines up with OpenAI's enterprise sales motion against Anthropic and Google.
The bar OpenAI has set is also a bar it now has to meet in the wild. A 52.5% reduction in medical, legal and financial hallucinations is the kind of number professional users will measure in production rather than on benchmarks, and the memory panel will draw scrutiny from privacy regulators already circling AI assistants that ingest email.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
