OpenAI updates its flagship model. On the program: a more flexible, more precise and more human model.
In the midst of a crisis amid a controversial agreement with the Pentagon, OpenAI is deploying GPT-5.3 Instant. The model, whose engine had already been deployed in Codex in February, is gradually replacing the GPT-5.2 range. Since the disappointing launch of GPT-5, the start-up has continued micro-updates to retain users who are increasingly courted by a Gemini which is gaining momentum and a Claude which is accumulating exclusive features at a sustained pace. The fact remains that GPT-5.3 Instant is not just a communication stunt, the model targets the small, very real irritants of ChatGPT.
A more flexible and precise model
GPT-5.3 Instant becomes the default model for text responses that don’t require reasoning. A GPT-5.3 Thinking version will arrive soon, promises OpenAI. The Instant version is much more permissive than the previous one. When confronted with sensitive questions, about health for example, the model tended not to respond. GPT-5.3 Instant erases this character trait and addresses these subjects more easily, with moderation but without morals.
Widely used as a search engine, GPT-5.3 Instant provides more flexibility during web searches. GPT-5.2 tended to rely primarily on sources from the web rather than its own knowledge, behaving almost more like a parrot than a real thinking tool. GPT-5.3 corrects this phenomenon: the model relies first on its internal knowledge and secondly uses recent web data to support its statements. The result? A model that should appear more intelligent, especially in an analysis situation.
OpenAI returns to the basics of generative AI
With GPT-5.3 Instant, OpenAI seems to be returning to the roots of what made it successful: a more intelligent and more human model. The tone becomes more natural and conversational, closer to GPT-4o if we read OpenAI’s intentions between the lines. The ready-made formulas and tics typical of generative models should be largely toned down. Finally, the model should be more consistent from one conversation to another, the user should have the impression of speaking to the same personality throughout their exchanges.
A direct illustration of this return to basics, ChatGPT should regain creative writing abilities. The teams specifically refined the model so that it was able to write text with precise tones, while maintaining stable consistency. We are therefore waiting for more human writing, closer, hopefully, to GPT-4.1. Finally, the model continues to reduce its hallucination rates, particularly in medicine, law and finance. “GPT‑5.3 Instant reduces hallucination rates by 26.8% when using the web and by 19.7% when relying solely on internal knowledge,” details OpenAI.
Stable pricing
On the training side OpenAI remains discreet. The system card does not mention any notable methodological changes compared to GPT-5.2. The fact remains that the model is visibly evolving in a sensitive area: the management of conversations related to mental health. Without making it a marketing argument, OpenAI has introduced new benchmarks on self-harm, emotional dependence and psychological distress. Issues that have been widely in the news in recent months.
On pricing, OpenAI maintains the prices of GPT-5.2 Instant: $1.75 input and $14.00 output, per million tokens. The model is available in ChatGPT and via the API under the identifier gpt-5.3-chat-latest. Please note, GPT-5.2-Instant will be disconnected on June 3.
GPT-5.3 Instant does not reinvent ChatGPT. But in a race where Gemini and Claude are gaining ground every week, even incremental progress counts. For OpenAI, the challenge is no longer to surprise: it is to not disappoint.
