GPT-53-Codex Released as OpenAI’s First AI Model to Assist in Its Own Development

GPT-5 was released by OpenAI as a . Thursday, a new artificial intelligence (AI) model for agentic coding tasks called 3-Codex was unveiled as one of the most recent models to be developed by s. But the new model will allow Codex to do more, including creating complex video games and apps from scratch, said San Francisco-based AI giant. Described as the company’s most capable agentic coding mode, OpenAI said that it can handle full workflows, debug entire codebases and research requirements while also deploy changes. Interestingly, the GPT-5 is an acronym for . Also the first AI model of an artificial intelligence giant, 3-Codex, is also the one that was a key part in the development of itself.

GPT-5.3-Codex Released

The AI giant detailed its latest frontier model of agentic coding for agentic encoding in a post. Today it is compatible with all of the global paid ChatGPT plans for mobile and desktop apps, command-line interface (CLI), integrated development environment (IDE) extension as well as web. In a short time, the company will provide the model via OpenAI’s application programming interface (API) and also through its own offering of open AI as well as an API.

The coding frontier performance of GPT-5 is combined with the new model’s ability to code effectively, according to . GPT-5 2-Codex with the reasoning and professional knowledge depth of 2- Codex. a single, ununified system of 2, making one and. This model has a 25 percent faster speed, so that it can tackle long-horizon projects with the focus on research, tool use and complex execution steps. Unlike earlier, users can steer it mid-task (which was not the case with s), and so are crucially capable of helping people to drive it. Consequently, users can receive progress updates, ask questions, suggest course corrections or debate approaches without the agent dropping context.

It even helped build itself up as a model, including . The Codex team was helped by early versions of the codex group in debugging training runs, managing deployment and diagnosing evaluation results. The team was impressed by how much acceleration this self-assistance offered during development, OpenAI says.

OpenAI shared benchmark scores based on internal tests of performance, as well as the term “” (along with its own evaluations) for performances. This reaches 56 on SWE-Bench Pro, an intense software engineering test across multiple languages that is hard to pass. 8 percent correct, excluding GPT-5. 4-CodeX at 56 2-Codex. GPT-5 – 4 percent and 5 percent of . 2 55 – 2. A total of six percent s are. Terminal-Bench 2. Jumping 0 jumps to the number of at least one , 77. 63 per cent from 64 percent. on 0 percent of the previous Codex variant. In OSWorld-Verified, which measures agent performance in visual desktop environments for productivity tasks, it reaches 64 percent. 7 percent vs 38 per cent for . For GPT-5, 2 percent for . 2 Codex (paraphrasing) 2-Codex.

GPT-5 3-Codex can create complex web games from unspecified instructions, generating autonomously more than millions of tokens. An official posting on the announcement post showed one demo of a racing game with maps, items and racers. It also makes more production-ready websites, automatically handling features such as discount displays or testimonial carousels. The full software lifecycle, which is supported outside pure coding, includes writing PRDs, editing copy, user research, building slide deck and analysis of spreadsheets and monitoring systems.

Similarly, OpenAI has focused on the safety guardrails of the AI model. PPT 5 – GPT-5. The first model OpenAI classifies as High capability under its Preparedness Framework for cybersecurity tasks is 3-Codex. This is an integrated cybersecurity security stack, which includes safety training and automated monitoring, trusted access controls as well as threat intelligence enforcement.