GPT-5.5 — a new model from OpenAI: what has changed and why it is needed

How GPT-5.5 differs from previous versions

In short, this is the first truly new model in a long time. All versions from GPT-5 to GPT-5.4 were updates built on top of the same underlying foundation. GPT-5.5 has been rebuilt from scratch.

Three key changes

Previously, ChatGPT could handle text, images, and audio, but these were processed by different internal models stitched together. GPT-5.5 processes all formats—text, images, audio, and video—within a single unified system.

The model was developed in collaboration with NVIDIA specifically for their servers. As a result, GPT-5.5 operates at the same speed as GPT-5.4, despite being significantly more powerful. Typically, smarter models run slower, but this limitation has been overcome here.

Before its launch, GPT-5.5 analyzed the operation of OpenAI’s servers and rewrote part of the code that manages workload distribution. Processing speed increased by more than 20%. The model improved the infrastructure that supports it.

What it is designed for

The primary goal of GPT-5.5 is not just to answer questions, but to perform tasks independently. You can give it a complex, poorly formulated multi-part task, and it will figure out what needs to be done, use the necessary tools, verify the results, and see the job through to completion.

OpenAI’s Chief Scientific Officer, Jakub Pachocki, noted that the company sees significant improvements in the short term and extremely significant ones in the medium term.

Mark Chen, Director of Research at OpenAI, added that the model shows noticeable gains in scientific and technical tasks and can genuinely help scientists move forward faster. He specifically highlighted the field of drug development.

How GPT-5.5 is already being used internally at OpenAI

OpenAI shared real examples of how company employees work with the model on a daily basis.

The finance department used GPT-5.5 to review nearly 25,000 tax documents—totaling over 71,000 pages. Previously, this work took two weeks longer.

One employee set up automated weekly report generation. This now takes much less time instead of the previous 5–10 hours.

A professor of immunology from Jackson Laboratory asked GPT-5.5 Pro to analyze a large dataset of genetic information. According to him, work that his team would have spent several months on was completed in a single session.

Today, more than 85% of OpenAI employees use the model every week. This includes not only developers but also marketers, finance professionals, communications specialists, and data analysts.

GPT-5.5 Instant — the new default model

Alongside GPT-5.5, OpenAI released another version: GPT-5.5 Instant. This is now the standard model for ChatGPT, replacing GPT-5.3 Instant.

The main difference with this version is that it makes fewer errors and hallucinates facts less frequently, especially in medicine, law, and finance. At the same time, it works just as fast as the previous version.

In tests, the model showed notable growth: in the AIME 2025 math test, it scored 81.2 compared to 65.4 for its predecessor. In the MMMU-Pro multimodal reasoning test, it scored 76.0 against 69.2.

A separate feature is that ChatGPT now remembers context. The model can look into your past conversations, uploaded files, and even Gmail to provide more accurate and personalized answers. However, you can see where each piece of information comes from and delete or update outdated data.

If you share a chat with someone, these sources will not be visible to the other person. Currently, this feature is available for Plus and Pro subscriptions in the browser; it will be available on mobile devices later.

For developers, the model is available in the API under the name chat-latest. The previous version, 5.3, will remain available for another three months.

Who has access and how much it costs

GPT-5.5 is available to paid ChatGPT users—Plus, Pro, Business, and Enterprise subscriptions. Free users may have limited or indirect access to the model through restricted modes and quotas.

The model is also open via API for developers and companies integrating ChatGPT into their products. The cost via API is $5 per million input tokens and $30 per million output tokens.

The GPT-5.5 Pro version is designed for the most complex and critical tasks where maximum accuracy is essential. It costs $30 per million input tokens and $180 per million output tokens. It is available only for Pro, Business, and Enterprise subscriptions.

For those who do not need an instant response, there is a Batch mode with a 50% discount. Requests are processed within 24 hours, but the price is the same as for GPT-5.4.

What people are saying about it

Developer reaction is generally positive, but with caveats. Developer Matt Schumer wrote that this is a “huge leap forward,” but for most users, the difference may be barely noticeable due to the already high level of previous models.

Engineers at NVIDIA, who received early access, described the results as “stunning” and “changing the approach to work.” According to NVIDIA, debugging code, which previously took days, is now completed in hours.

According to OpenAI President Greg Brockman, he called GPT-5.5 “a new level of intelligence for real-world work.”

He explained that the model can:
— look at an ambiguous task
— understand exactly what needs to be done next
— independently construct the steps to solve it.

He also gave an example: one mathematics professor was able to create an application for algebraic geometry in just 11 minutes by simply giving the model a single prompt.

Dan Shipper, founder and CEO of Every, described GPT-5.5 as “the first programming model I’ve used with a truly clear conceptual understanding.”

He described an experiment where, after a complex system failure, he partially rolled back changes and checked if the model could reproduce the engineering solution the team had arrived at. According to him, GPT-5.4 failed this task, whereas GPT-5.5 was able to restore the logic of the fix by relying on the “broken” state of the system.

Pietro Schirano, CEO of MagicPath, noted a similar impression: “It honestly feels like I’m working with a higher intelligence, and a sense of respect almost arises.”

He described a case where the model consolidated major changes in the frontend and refactoring that simultaneously affected the main project branch. According to him, GPT-5.5 handled combining these complex changes in about 20 minutes—a task that usually requires significant manual involvement from the team.