OpenAI releases GPT-5.5, moving ChatGPT closer to an AI super app

OpenAI has released GPT-5.5 and presents it not as another incremental ChatGPT update but as part of a larger strategy: turning its products into a single work interface where the model does more than answer questions. It walks through the steps of a task itself, uses tools and sees the work through to a result. For the market, this is an important signal: the race among AI assistants is shifting from the quality of individual answers to the quality of autonomous execution of real tasks on a computer.

The announcement came on April 23, 2026. The company calls GPT-5.5 its smartest and most intuitive model release to date. According to OpenAI president Greg Brockman, the new version brings the company closer to a more agentic and natural way of working with a computer. This is not just a chatbot but a system that can be given a vague, multistep assignment, such as writing and debugging code, gathering material on the web, analyzing data, or preparing a document or spreadsheet, and that returns not only text but a finished result.

According to OpenAI's official materials, GPT-5.5 keeps latency at the level of GPT-5.4 but solves the same tasks with fewer tokens and performs better on specialized tests. On Terminal-Bench 2.0, which tests complex command-line work, the model scored 82.7% versus 75.1% for GPT-5.4. On OSWorld-Verified, which assesses the ability to work independently in a computer environment, it scored 78.7%. On GDPval, which measures the quality of formalized knowledge work across dozens of professions, it scored 84.9%. OpenAI also claims GPT-5.5 outperforms earlier models and, in a number of scenarios, beats strong competitors such as Claude Opus 4.7 and Gemini 3.1 Pro.

The main emphasis is on two classes of tasks: agentic development and intellectual office work. In Codex, the model is supposed to hold the context of a large project better, find the causes of failures, plan fixes and check its own work with tools. In ordinary work scenarios, it should gather information, build spreadsheets and presentations, and help with analytics, finance and research tasks. OpenAI cites internal usage examples: more than 85% of the company's employees already use Codex weekly, and some teams automate reports, request processing and large volumes of document work.

This logic ties directly into the AI super app idea: in a corporate post dated April 8, OpenAI described a single interface that would combine ChatGPT, Codex, agentic web browsing and actions in other tools.

Another important area is scientific and technical research. The company says GPT-5.5 has become noticeably stronger on tasks that require more than producing an answer, going through a full cycle: collecting data, testing hypotheses, interpreting results and proposing the next step. Examples include improvements on GeneBench and BixBench, as well as cases where the model helped with biological data analysis and mathematical proofs.

At the same time, OpenAI tightened restrictions around cybersecurity and biological risks: the model went through additional testing, and for sensitive scenarios the company introduced stricter classifiers and access modes.

On April 23, GPT-5.5 began rolling out to Plus, Pro, Business and Enterprise users in ChatGPT and Codex, and on April 24 OpenAI clarified that GPT-5.5 and GPT-5.5 Pro were also available through the API.

What does this mean? OpenAI looks less and less like the developer of a standalone chatbot and more like a provider of a universal work environment, where the model takes on not only text generation but part of the execution. If GPT-5.5 really performs reliably over long chains of actions, the market will get a new standard: models will be compared not by polished demos and single answers, but by how reliably they can carry a complex task through to the end without constant manual supervision.