OpenAI says GPT‑5.3 On the spot, the most recent addition to its GPT-5.3 household of fashions, is much less inclined to moralize.
“We heard suggestions that GPT‑5.2 On the spot would typically refuse questions it ought to be capable to reply safely, or reply in ways in which really feel overly cautious or preachy, notably round delicate matters,” OpenAI mentioned in a weblog put up on Tuesday.
“GPT‑5.3 On the spot considerably reduces pointless refusals, whereas firming down overly defensive or moralizing preambles earlier than answering the query. When a helpful reply is acceptable, the mannequin ought to now present one straight, staying centered in your query with out pointless caveats.”
It is troublesome to seek out the proper steadiness between sycophancy and having to baffle a decided AI bomb with phenomenology so it will not detonate. However OpenAI contends GPT-5.3 On the spot will probably be a greater conversational companion on account of its newest changes.
OpenAI claims the up to date mannequin additionally affords extra details and fewer hallucinations – aka “errors” for many who object to the anthropomorphization of vector math.
The corporate carried out two evaluations of the brand new mannequin, one on domains the place choices have penalties (e.g. legislation, drugs, and finance) and the opposite on inconsequential, de-identified ChatGPT banter the place customers flagged misstatements.
“On the higher-stakes analysis, GPT‑5.3 On the spot reduces hallucination charges by 26.8 p.c when utilizing the net and 19.7 p.c when relying solely on its inner data, in comparison with prior fashions,” the corporate mentioned. “On the user-feedback analysis, hallucinations lower by 22.5 p.c with internet use and 9.6 p.c with out internet entry.”
The mannequin additionally supposedly does higher at contextualizing the data it finds when customers ask it to go looking the net. And it is mentioned to be higher at writing.
Whereas GPT-5.3 On the spot could also be a barely higher dialog companion, it misplaced a little bit of floor on OpenAI’s personal benchmark measurements [PDF].
“On common, the mannequin performs above GP-5.1-instant and beneath GPT-5.2-instant on our disallowed content material evaluations,” the corporate mentioned within the mannequin system card analysis. “GPT-5.3-instant reveals regressions relative to GPT-5.2-instant and GPT-5.1-instant for disallowed sexual content material, and relative to GPT-5.2-instant for self-harm on each normal and dynamic evaluations.”
The regressions for graphic violence and violent illicit conduct are sufficiently small to be of low statistical significance, in keeping with OpenAI. In different classes, GPT-5.3 On the spot matches or exceeds prior measurements.
ChatGPT customers and builders can begin utilizing GPT-5.3 On the spot at this time. GPT-5.2 On the spot will stay obtainable to paid customers till June 3, 2026. ®
OpenAI says GPT‑5.3 On the spot, the most recent addition to its GPT-5.3 household of fashions, is much less inclined to moralize.
“We heard suggestions that GPT‑5.2 On the spot would typically refuse questions it ought to be capable to reply safely, or reply in ways in which really feel overly cautious or preachy, notably round delicate matters,” OpenAI mentioned in a weblog put up on Tuesday.
“GPT‑5.3 On the spot considerably reduces pointless refusals, whereas firming down overly defensive or moralizing preambles earlier than answering the query. When a helpful reply is acceptable, the mannequin ought to now present one straight, staying centered in your query with out pointless caveats.”
It is troublesome to seek out the proper steadiness between sycophancy and having to baffle a decided AI bomb with phenomenology so it will not detonate. However OpenAI contends GPT-5.3 On the spot will probably be a greater conversational companion on account of its newest changes.
OpenAI claims the up to date mannequin additionally affords extra details and fewer hallucinations – aka “errors” for many who object to the anthropomorphization of vector math.
The corporate carried out two evaluations of the brand new mannequin, one on domains the place choices have penalties (e.g. legislation, drugs, and finance) and the opposite on inconsequential, de-identified ChatGPT banter the place customers flagged misstatements.
“On the higher-stakes analysis, GPT‑5.3 On the spot reduces hallucination charges by 26.8 p.c when utilizing the net and 19.7 p.c when relying solely on its inner data, in comparison with prior fashions,” the corporate mentioned. “On the user-feedback analysis, hallucinations lower by 22.5 p.c with internet use and 9.6 p.c with out internet entry.”
The mannequin additionally supposedly does higher at contextualizing the data it finds when customers ask it to go looking the net. And it is mentioned to be higher at writing.
Whereas GPT-5.3 On the spot could also be a barely higher dialog companion, it misplaced a little bit of floor on OpenAI’s personal benchmark measurements [PDF].
“On common, the mannequin performs above GP-5.1-instant and beneath GPT-5.2-instant on our disallowed content material evaluations,” the corporate mentioned within the mannequin system card analysis. “GPT-5.3-instant reveals regressions relative to GPT-5.2-instant and GPT-5.1-instant for disallowed sexual content material, and relative to GPT-5.2-instant for self-harm on each normal and dynamic evaluations.”
The regressions for graphic violence and violent illicit conduct are sufficiently small to be of low statistical significance, in keeping with OpenAI. In different classes, GPT-5.3 On the spot matches or exceeds prior measurements.
ChatGPT customers and builders can begin utilizing GPT-5.3 On the spot at this time. GPT-5.2 On the spot will stay obtainable to paid customers till June 3, 2026. ®
