OpenAI has released GPT-5.3 Instant, the default model powering ChatGPT for most users, with a primary focus on reducing factual errors rather than improving raw speed. The company says the update cuts hallucinations by up to 26.8% compared to its predecessor.
The model is available immediately on both ChatGPT and the API. OpenAI confirmed it is working on extending the 5.3 update to its Thinking and Pro model tiers, though no timeline was given.
What the Numbers Actually Show
OpenAI ran two separate internal evaluations to measure the update’s accuracy gains. One tested higher-stakes domains including medicine, finance, and law. The other drew on user feedback data.
The results broke down as follows:
- 26.8% reduction in hallucinations when using web search, based on internal higher-stakes evaluations
- 19.7% improvement in reliability when the model relies on its own internal knowledge
- 22.5% decrease in hallucinations on web-assisted queries, based on user feedback
The gains come primarily from how the model balances live web results with its own training. According to OpenAI, the previous version, GPT-5.2 Instant, had a tendency to overindex on search results, often returning long link lists or loosely connected information rather than synthesized answers.
Less Refusals, Different Tone
“This update focuses on the parts of the ChatGPT experience people feel every day: tone, relevance, and conversational flow. These are nuanced problems that don’t always show up in benchmarks, but shape whether ChatGPT feels helpful or frustrating,” the company said in a blog post.
GPT-5.3 Instant also significantly reduces refusals. OpenAI acknowledged that GPT-5.2 frequently declined to answer questions that fell within acceptable guidelines, and sometimes responded with what it described as overly cautious or preachy framing around sensitive topics. The new model is designed to answer directly, without defensive preambles or unsolicited moral caveats.
OpenAI also said it intends to keep the model’s conversational personality more consistent across future updates, so users do not encounter noticeable tonal shifts between versions.
Limitations Remain
The model still has notable gaps. Responses in Korean and Japanese remain stilted. Support for adult content is absent, with an OpenAI spokesperson confirming the company is still determining how to balance user freedom with its safety standards, and no release date has been set for that capability.
On the safety front, OpenAI’s own benchmarking showed GPT-5.3 Instant performed well against disallowed content but did not match the safety scores of GPT-5.2 Instant. The company noted those results may shift after the model is deployed at scale.
A Crowded Accuracy Race
The release reflects a broader competitive dynamic. Anthropic recently claimed its Claude Sonnet 4.6 produces fewer hallucinations. Google pulled its Gemma 3 model after it generated false information about a lawmaker. Accuracy has become the visible battleground, replacing the earlier emphasis on speed and benchmark scores.
OpenAI’s framing of GPT-5.3 Instant as a reliability-first release signals a deliberate repositioning, particularly aimed at enterprise customers who need models that adhere to factual information in consequential domains.
Photo by Rolf van Root on Unsplash
This article is a curated summary based on third-party sources. Source: Read the original article
