If you think that vibe coding can no longer be impressive, then OpenAi enters the stage with GPT-5 Codex and says: “Hold my scripts.”
On Monday, OpenAI officially released GPT-5 Codex, a specialized version of its flagship LLM, which was designed for autonomous programming tasks. This is not another ordinary assistant to the code, which can only complete, but a real beast that is able to work on complex programming problems for several hours without a break. How does GPT-5 Codex work from OpenAI?
Coding (and thinking) on demand – from a second to hours
The most interesting feature of GPT-5 Codex is its ability to dynamically adapt the “thinking time” depending on the complexity of the task. Unlike previous AI tools, which had a predetermined reflection time, this model can decide during work: “Hmm, it is more complicated than I thought – I need more time.” It is worth adding that in the last few months the flagship models of LLM have been able to think longer about solving the problem-these options are available both in Grok 4, Gemini 2.5 or ChatgPT-5.
Alexander Embiricos, head of the Codex product at Openai, explained this in a way that sounds almost like science fiction. Codex may say after five minutes that he needs to spend an hour more. During internal tests, OPENAI observed cases when GPT-5 Codex himself worked on large-scale tasks requiring re-invoicing for over seven hours, items on implementation, repairing errors in tests and providing effective solutions. Of course, I would approach the 7-hour vibe coding sessions served by Codex, but the current possibilities achieved by “AI thinking” are optimistic, especially in autonomous coding.
This solves a key limitation of earlier assistants for coding – a constant need for human intervention. Now you can throw him a complicated task before going to bed and wake up with a ready solution. Sounds like a dream of every developer, right?
What is Vibe Coding?
Although coding is getting louder and louder about this new method, it is worth introducing what Vibe Coding really is. According to the definition, this is an approach to programming, in which not only the code itself plays a key role, but also the atmosphere, energy and creative “flow” accompanying its creation. Of course, the basis of vibe coding is the use of Genai tools.
It is combining technology with intuition and aesthetics – instead of focusing only on pure logic, the programmer follows the rhythm, mood and emotions that give the project a unique character. Thanks to this, the code becomes not only a tool, but also a form of expression in which the vibe of the whole process is important. Importantly, the vast majority of the most important tasks are performed by LLMs, thanks to which Vibe Coding is also available to people who are just starting the programming adventure.
Code review on steroids
GPT-5 Codex is not only a code writing machine, but above all an expert on its quality control. The model has undergone specialized training in conducting comprehensive code reviews, which distinguishes it from competitors focusing mainly on auto -complete.
Codex can move around all code databases, analyze dependencies and perform tests to verify correctness. During the assessments made by experienced software engineers, GPT-5 Codex from OPENAI showed much less erroneous comments and provided more “high-impact comments” compared to previous versions.
Aaron Wang, senior software engineer in Duolingo, did not hide his enthusiasm:
Codex exceeded our backend code review (and was) the only tool that identified complex problems with reverse compatibility
This means that AI not only writes better code than some people – now it also reviews it better.
Does OpenAI begin the war for the kingdom of coding?
The introduction of GPT-5 Codex is a clear declaration of war on the increasingly competitive AI tools for programming tools. Cursor from Anysphere recently boasted $ 500 million annual repetitive revenues, while GitHub Copilot from Microsoft still dominates in the daily work of developers. Meanwhile, Windsurf became such a “hot commodity” that both Google and Cognition tried to take over.
Opeli does not intend to be behind. GPT-5 Codex exceeds the standard GPT-5 model on Swe-Bench Verified (benchmark measuring agent coding skills) and in the tasks of re-invoicing code from large repositories. The model has been designed for trouble -free work in various programming environments: terminals, IDE, Internet browsers, githuba and mobile devices.
Availability and integration
GPT-5 CODEX is now available on all Codex platforms, including in command line interfaces, integrated programming environments, integration with GitHub and in mobile applications for ChatgPT Plus subscribers, Pro, Business, Edu and Enterprise subscribers.
This interplalatform integration allows developers to transfer work between local environments and cloud agents without losing context, thus solving one of the biggest ills in workflows in the development of software development supported by AI.
Performance that impresses
The OpenAI coding agent has been significantly improved. Its performance has been improved even about 90% For some cloud -based tasks, shortening the time of performing new tasks and follow -up activities. The model can now create websites, provide useful and correct comments to the code, and conduct in -depth code reviews for errors.
Safety primarily
GPT-5 Codex performs all its work in the Sandbox environment. Access to the network and other tools is turned off by default, but it can be activated by developers according to their needs. After accessing access, the agent can surf the internet, build and iterate projects.
However, due to the ability to perform tasks and inherent susceptibility to burglary, AI can be manipulated to perform unwanted activities. As a result, OPENAI assessed GPT-5 Codex as a high risk in biological and chemical domains, activating additional protection to detect and counteract its use to hurt people. This is particularly important, considering how many users try their hand at jailbreaking LLM models.
Codex from OPENAI COMMENT CODING forever?
GPT-5 Codex is a significant change in the world of programming tools supported by AI. The ability to work for hours, autonomous work on complex programming problems, combined with advanced Code Review capabilities and interpelform integration, can fundamentally change the way professionals and amateurs create software.
The question is: are we ready for the world in which AI not only helps us cod, but actually takes over a significant part of this work?
