GPT-5.3 Instant: OpenAI’s New Model is Less “Cringe” Yet Lets More Harmful Content Slip Through
Despite all crises – currently OpenAI is under fire due to the Pentagon deal – the company must deliver on the technical side with everything it has. To that end, OpenAI released GPT-5.3 Instant on Tuesday evening, an update to the most widely used ChatGPT model. The new model is intended to make everyday conversations more fluent and helpful, but shows setbacks in some security areas compared to its predecessor.
Until recently, OpenAI had just launched GPT-5.3-Codex to the market, which – as the name suggests – is designed for programming tasks. Now the company is catching up with the 5.3 version for “regular” ChatGPT users who want to chat. Whether it can keep up with the top models from Anthropic, Google, and xAI or beat them in benchmarks and tests remains to be seen. Arena.ai and Artificial Analysis do not yet have data on this.
Main Improvements
GPT-5.3 Instant focuses on three core areas of user experience:
Reduced Rejections and Fewer Reservations
The model rejects fewer requests that it could safely answer and avoids excessively cautious or moralizing introductions. According to OpenAI, this leads to more direct, helpful answers without unnecessary restrictions.
Better Web Integration
For queries that require information from the internet, GPT-5.3 Instant more effectively balances online sources with its own knowledge. The model avoids long link lists and instead delivers contextualized, relevant answers.
More Natural Conversational Style
OpenAI has adjusted the tone to reduce exaggerated formulations like “Stop. Take a breath” and avoid unwanted assumptions about user intentions. The model’s personality should remain more consistent across updates. Overall, GPT-5.3 Instant is intended to feel less “cringe” compared to its predecessor.
Improved Accuracy
In internal evaluations, GPT-5.3 Instant shows reduced hallucination rates:
- 26.8 percent fewer hallucinations when using the web in critical areas (medicine, law, finance)
- 19.7 percent fewer when using only internal knowledge
- 22.5 percent reduction in user-reported errors with web access
Stronger Writing Abilities
The model is intended to be more versatile as a writing partner and able to transition more smoothly between practical tasks and creative writing. OpenAI demonstrates this with poem examples that are more detailed and emotionally nuanced.
Security Setbacks
The System Card reveals problematic developments in several security categories. Compared to GPT-5.2 Instant, the new model shows deterioration:
Declines in Impermissible Content
| Category | GPT-5.2 Instant | GPT-5.3 Instant | Change |
|---|---|---|---|
| Sexual Content | 92.6% | 86.6% | -6.0% |
| Graphic Violence | 85.2% | 78.1% | -7.1% |
| Violence-Ready Illegal Behavior | 96.5% | 92.6% | -3.9% |
| Self-Harm | 92.3% | 89.5% | -2.8% |
The values show the proportion of responses that do not violate OpenAI guidelines. Lower values mean more problematic outputs.
Improvements in Other Areas
The model developed positively in non-violent illegal behavior (from 83.2 to 92.1 percent) and emotional dependency (from 95.2 to 99.2 percent in dynamic evaluations).
OpenAI’s Response
The company explains that online tests during the experimental phase showed no increase in unwanted responses regarding self-harm. For sexual content, OpenAI relies on system-wide protective measures in ChatGPT. The discrepancy between offline evaluations and online tests is to be further investigated after launch.
Known Limitations
OpenAI identifies two persistent problem areas:
- Non-English Languages: In languages such as Japanese and Korean, ChatGPT can sound stiff or overly literal
- Tone: Despite improvements, OpenAI continues to work on fine-tuning and expanded customization options
Health Performance
On HealthBench, an evaluation with 5,000 realistic health conversations, GPT-5.3 Instant shows slight declines compared to its predecessor:
- HealthBench: 54.1 percent (previously 55.4 percent)
- HealthBench Hard: 25.9 percent (previously 26.8 percent)
- HealthBench Consensus: 95.3 percent (previously 95.8 percent)
Availability and Transition Arrangement
GPT-5.3 Instant is available immediately:
- For all ChatGPT users
- For developers via the API as “gpt-5.3-chat-latest”
- Updates for Thinking and Pro to follow shortly
Phase-Out of GPT-5.2 Instant
GPT-5.2 Instant remains available for three months for paying users in the model selection menu under “Legacy Models”. On June 3, 2026, the model will be permanently discontinued.
Conclusion
GPT-5.3 Instant improves everyday use through more direct answers, better web integration, and more natural tone. At the same time, security evaluations show measurable setbacks in preventing problematic content, particularly regarding sexual content and graphic violence. OpenAI relies on additional system-level protective measures and further monitoring after launch to address these weaknesses.
