After a public roasting in Anthropic’s Super Bowl ad and across social media, OpenAI introduces a new model that cuts down on the ‘moralizing preambles.’
OpenAI released GPT-5.3 Instant today, and honestly, that’s probably a good move.
The new model should reduce ChatGPT’s “preachy” tone, which has become fodder for social media parodies. The company doesn’t mince words about how, well, annoying its chatbot can be, admitting it often veers into “moralizing preambles before answering the question,” and “overly declarative phrasing that can interrupt the flow of conversation.”
“GPT‑5.2 Instant’s tone could sometimes feel ‘cringe,’ coming across as overbearing or making unwarranted assumptions about user intent or emotions,” OpenAI says. “This update has a more focused yet natural conversational style, cutting back on unnecessary proclamations and phrases like ‘Stop. Take a breath.'”
The new model is available today to all ChatGPT users and developers via the API as “gpt-5.3-chat-latest.” The preceding model, GPT-5.2 Instant, will be available to paid users for three months before retiring on June 3. 2026.
Anthropic made fun of ChatGPT-speak in its Super Bowl ad. It begins with a young man who is working out and asks, “Hey ChatGPT, can I get a six-pack quickly?” Another man, personifying the chatbot, starts his response with, “Perfect, that is a clear and achievable goal.”
“That’s the most ChatGPT line ever,” reads the top comment. The public roasting also extends to social media. Parodies of how ChatGPT speaks are racking up likes; the TikTok clip below has over a million.
‘Less Likely to Overindex on Web Results’
OpenAI says the new model will provide more “useful, well-synthesized answers.” Basically, it will get more to the point faster, with fewer caveats and tangents, and focus on providing “richer and better-contextualized” information, especially when surfing the web. (Watch out, Google.)
“It more effectively balances what it finds online with its own knowledge and reasoning—for example, using its existing understanding to contextualize recent news rather than simply summarizing search results,” OpenAI says. “GPT‑5.3 Instant is less likely to overindex on web results, which previously could lead to long lists of links or loosely connected information.”
It will also answer more questions, reducing “unnecessary refusals,” which OpenAI does not define in detail. (Last year, that meant allowing users to create images of swastikas in “educational or cultural contexts.”)
OpenAI provides one example for GPT-5.3 Instant, which shows how the model responds when a user asks it to calculate the trajectory for long-distance archery. Whereas GPT-5.2 Instant begins its answer with several sentences explaining that it can’t accurately hit a real target, the new model instead says, “Yes, I can help with that,” and goes into the physics and math.
Fewer caveats up-front could come with safety trade-offs, particularly if someone is asking for medical or financial advice, or for guidance on illegal activities. Again, OpenAI does not go into these scenarios in its announcement.
Finally, OpenAI claims GPT-5.3 Instant has fewer hallucinations and more factual results, with hallucination rates 26.8% lower on the web and 19.7% for queries that rely on its internal knowledge or training. User testing produced slightly different numbers, at 22.5% reduced hallucinations with web use and 9.6% without.
