Introduction to OpenAI’s GPT-5.5 Instant Model

OpenAI has recently launched its latest model, GPT-5.5 Instant, superseding the previous GPT-5.3 Instant, which was released on March 3. This upgrade emphasizes greater accuracy and concise responses over the model’s personality traits. The latest iteration aims to deliver reliable information across critical domains.

Impressive Reductions in Hallucinations

In internal assessments, GPT-5.5 Instant demonstrated a remarkable 52.5% reduction in hallucinated claims when responding to high-stakes prompts in fields such as medicine, law, and finance. Furthermore, it achieved a 37.3% decrease in inaccuracies in conversations that users had previously flagged for errors, marking a significant step in ensuring factual reliability.

Enhanced Performance in Standardized Tests

The advancements in GPT-5.5 Instant are also evident in its performance on standardized tests. The model scored an impressive 81.2 on the AIME 2025 math test, a significant increase from the 65.4 score of its predecessor. Additionally, it achieved 76 on the MMMU-Pro multimodal reasoning benchmark, compared to 69.2 of GPT-5.3 Instant, showcasing its enhanced reasoning capabilities.

Introduction of Memory Source Visibility

To provide more transparency in how it generates answers, OpenAI is updating ChatGPT to include memory source visibility. This feature allows users to see which past interactions, files, or Gmail data influenced a specific response, promoting a better understanding of the model’s decision-making process.

User Control Over Data and Sources

Users will have the ability to delete outdated sources or rectify inaccuracies, empowering them with greater control over their interactions with the model. The memory source feature is initially available to Plus and Pro subscribers on the web, with plans for mobile access rolling out soon. Free, Go, Business, and Enterprise users can expect this feature in the upcoming weeks.

Data Security and Shared Chats

OpenAI has also ensured that shared chats will not reveal memory sources to recipients, prioritizing user privacy and data security. This measure is crucial for maintaining trust, especially in sensitive discussions. Developers can access the GPT-5.5 model through the API, designated as “chat-latest,” while the previous GPT-5.3 Instant will remain available to paid users for a transition period of three months.

Transitioning from Previous Models

Understanding that the retirement of models can lead to user dissatisfaction, OpenAI acknowledges past experiences. For example, when GPT-4o was deprecated in February 2026, many users expressed their attachment by signing petitions to sustain its use. In light of this, OpenAI is providing a three-month transition period this time, allowing users to adapt seamlessly to the new model.

OpenAI Unveils New Model: GPT-5.5

OpenAI has recently upgraded its default AI model to GPT-5.5, marking a step forward from the previous version, GPT-5.3 Instant. This latest iteration underscores a commitment to accuracy and clarity, prioritizing reliable information over personality traits that may have characterized earlier versions.

Enhanced Accuracy in High-Stakes Scenarios

In internal assessments, GPT-5.5 exhibited a remarkable 52.5% reduction in hallucinated statements when tasked with critical prompts related to sectors such as medicine, law, and finance. Furthermore, during user-generated conversations flagged as containing factual inaccuracies, the model achieved a 37.3% reduction in erroneous claims.

Improved Performance on Benchmark Tests

The new model not only excels in accuracy but also posts significant gains in benchmark tests. For example, GPT-5.5 scored an impressive 81.2 on the AIME 2025 math exam, a notable improvement from the 65.4 score of its predecessor. Additionally, it outperformed on the MMMU-Pro multimodal reasoning benchmark with a score of 76, up from 69.2.

Commitment to Transparency with Memory Source Visibility

In tandem with the model upgrade, OpenAI is enhancing transparency for users regarding how the AI generates its responses. A new feature will show users the sources of information, including past conversations and files, that informed any given reply. This transparency aims to bolster user trust and understanding of AI-generated content.

User Control Over Information

With the updated memory source feature, users have the capability to delete outdated information or make corrections to inaccuracies. This control will initially be available to Plus and Pro subscribers on the web, with mobile access being introduced shortly thereafter. Free, Go, Business, and Enterprise users will have this functionality available in the following weeks.

Shared Chats and Developer Access

Importantly, shared chats will not disclose memory sources to recipients, ensuring user privacy. For developers, GPT-5.5 is accessible via the API under the label “chat-latest,” while the previous model remains available to paid users for a three-month window before it is phased out.

Transitioning Users Towards the Future

OpenAI understands that transitioning from older models can lead to user concerns. In response to feedback from previous model retirements, the company is providing a three-month period for users to adapt to GPT-5.5, ensuring a smoother transition into this new phase of AI interaction.