The big picture: As the race for AI supremacy intensifies, both OpenAI and Anthropic unveiled upgraded models this week. Anthropic's Claude Opus 4.6 marks a significant evolution in how AI tackles complex workplace tasks while OpenAI's GPT-5.3 Codex is the company's most advanced agentic coding model, capable of contributing to its own development.
GPT-5.3 Codex merges the advanced coding capabilities of GPT-5.2 Codex with the reasoning and professional knowledge of GPT-5.2 into a single, unified model that is 25 percent faster than its predecessors. According to OpenAI, the model even contributed to its own development, as early versions were used to debug training processes, manage deployment, and analyze test results.
The release represents a major leap for Codex, evolving it from an agent capable of writing and reviewing basic code into one of the most powerful coding models available – capable of performing nearly any task a developer can handle. It is also significantly faster than previous versions, achieving record-breaking scores on SWE-Bench Pro and Terminal-Bench.
To test GPT-5.3 Codex's web development skills and long-running agentic capabilities, OpenAI had the model create two "highly functional (and) complex" games. The first is an upgraded version of an existing racing title, Voxel Velocity, while the second is a brand-new diving game, Dive In, which lets players explore reefs, collect fish, and navigate underwater hazards.
To build the games, the Codex team leveraged the model's recently introduced "develop web game" skill in the Mac app. The team provided pre-selected, generic follow-up prompts, such as "fix the bug" and "improve the game." GPT-5.3 Codex reportedly executed these tasks autonomously over the course of several days, producing the desired results without human intervention.
Both games are currently playable on OpenAI's official website, offering users a firsthand look at GPT-5.3 Codex's capabilities. The release also highlights the potential for fully self-generating AI, raising concerns about job displacement in the tech sector and accelerating a future in which human programmers could become increasingly redundant.
Beyond creating fully functional video games, GPT-5.3 Codex can also produce polished, ready-to-publish websites. The model is designed to assist coders, designers, product managers, and data scientists with a wide range of software tasks, including debugging, deploying, monitoring, writing PRDs, editing copy, conducting user research, running tests, analyzing metrics, and more.
Trained on and deployed using Nvidia's GB200 NVL72 systems, GPT-5.3 Codex is now available to paid ChatGPT subscribers across all platforms – including web, mobile app, CLI, and IDE extension – with API access expected to launch soon.
GPT 5.3 Codex, OpenAI's new agentic coding model, helped create itself
