OpenAI's GPT-5.3 Instant Aims to Be Less Preachy While Walking Back Defense Department Deal

OpenAI has unveiled GPT-5.3 Instant, the latest iteration of its language model family, promising users a less "preachy" experience that cuts down on unnecessary refusals and moralizing preambles. The AI giant says the updated model will provide more direct answers to questions without the defensive caveats that characterized previous versions, particularly around sensitive topics.

"We heard feedback that GPT-5.2 Instant would sometimes refuse questions it should be able to answer safely, or respond in ways that feel overly cautious or preachy, particularly around sensitive topics," OpenAI stated in a Tuesday blog post announcing the release.

The timing of this announcement is particularly interesting given the company's recent controversy over its Defense Department contract. Just weeks ago, OpenAI celebrated a deal to provide AI services to the DoD, a move that drew sharp criticism from competitors and users alike. Anthropic, OpenAI's chief rival, had recently rejected similar overtures from the Pentagon, refusing to remove model guardrails for autonomous weapon development and citizen surveillance.

Anthropic's principled stand appears to have resonated with users. A company spokesperson revealed that Claude has become the most downloaded app on Google Play in the US, with Monday, March 2 marking the largest single day ever for signups. The AI assistant also claimed the top spot as the most downloaded free app in the iOS App Store.

In response to the backlash, OpenAI CEO Sam Altman announced plans to amend the initial DoD contract "to prohibit deliberate tracking, surveillance, or monitoring of US persons or nationals, including through the procurement or use of commercially acquired personal or identifiable information." Altman also issued a mea culpa of sorts, acknowledging that the company's quick pivot to the DoD deal "looked opportunistic and sloppy" and was an attempt to "de-escalate things and avoid a much worse outcome."

On the technical front, GPT-5.3 Instant brings several improvements beyond just being less moralistic. OpenAI claims the model significantly reduces hallucination rates - what the company calls "mistakes" for those who object to anthropomorphizing vector math. According to company evaluations, hallucination rates dropped by 26.8 percent when using web search and 19.7 percent when relying on internal knowledge, compared to prior models.

The updated model also supposedly performs better at contextualizing information found during web searches and shows improvements in writing capabilities. However, the gains aren't universal. On OpenAI's own benchmark measurements, GPT-5.3 Instant performs above GPT-5.1-instant but below GPT-5.2-instant in certain categories, particularly around disallowed content.

Specifically, the model shows regressions for disallowed sexual content and self-harm content on both standard and dynamic evaluations. The company notes that regressions for graphic violence and violent illicit behavior are small enough to be of low statistical significance.

ChatGPT users and developers can start using GPT-5.3 Instant immediately, though GPT-5.2 Instant will remain available to paid users until June 3, 2026. The release comes as OpenAI continues to navigate the delicate balance between providing useful AI assistance and maintaining appropriate safeguards, all while trying to rebuild trust with a user base that increasingly values principled stances over rapid expansion into controversial partnerships.