GPT-5.3 Instant update: improvements and changes to the ChatGPT model

GPT-5.3 Instant has been released — it's an upgrade of the main ChatGPT AI model.

This is not a global update, but purely cosmetic on top of the old model, based on user feedback.

I've analyzed the main points.

1. Works better with texts.

OpenAI claims that the neural network now generates text better (I will definitely check this).

The model better maintains structure, adds details more accurately, and sounds more natural.

2. Fewer glitches.

The company provided internal accuracy assessments.

Error reduction:
–26.81% when working with web search
–19.71% when responding based only on internal knowledge

3. Fewer refusals.

In GPT-5.2, the neural network sometimes refused to answer even permissible questions or started the response with long warnings.

In GPT-5.3, such situations will be less frequent. If the request is permissible, the AI responds to the point, without long disclaimers.

4. Improved work with web search.

GPT-5.3 better aligns the found information with its knowledge, rather than just retelling the search results.

Plus, the model less frequently overloads answers with lists of sources and tries to bring the main point to the forefront.

5. Calm tone.

In GPT-5.2, answers sometimes began with unnecessary emotional phrases like «Stop. Take a breath».

In GPT-5.3, such inserts were removed, and the answers became more neutral.

Conclusion. I will highlight two points:

Firstly, OpenAI rarely mentions text generation in updates. Here, they directly state that the model writes better — and it's good that this direction is not forgotten.

Secondly, the upgrade was made after feedback from users, focusing on practice and real tasks, rather than ephemeral benchmark scores.

P.S. It's important to note that this release is specifically for the fast Instant model. Updates for other modes: Thinking and Pro will be released later.