ChatGPT just got Updated

OpenAI's GPT 5.3 Codeex introduces significant speed improvements and enhanced agentic coding capabilities, marking a major leap towards autonomous self-improvement and broader professional applications.

• GPT 5.3 Codeex offers a 25% speed increase and is moving towards autonomous self-improvement in coding.

• The model can autonomously generate complex applications and excels at understanding underspecified prompts.

• Codeex 5.3 significantly improves computer control and general knowledge work capabilities, challenging industry rivals.

OpenAI has launched GPT 5.3 Codeex, a new model focused on agentic coding and capable of autonomous self-improvement. It boasts a 25% speed increase achieved by generating the same results with significantly fewer tokens. This update expands its utility beyond coding to general knowledge work, directly competing with Anthropic's Opus 4.6.

Agentic Coding and Autonomy

• 00:01:27 GPT 5.3 Codeex represents a move towards autonomous self-improvement, with previous versions instrumental in creating the current iteration by debugging training, managing deployment, and diagnosing results. This allows Codeex to function as an agent capable of performing nearly any task a developer or professional can on a computer, indicating a major industry shift towards long-horizon tasks and agent teams.

Performance Enhancements

• 00:01:58 The 25% speed increase in GPT 5.3 Codeex is primarily due to its ability to achieve comparable results using significantly fewer output tokens, as demonstrated by benchmarks like SweetBench Pro. For example, it uses 43,000 tokens compared to 91,000 for its predecessor, GPT 5.2 Codeex, while also showing a 10+ point bump in TerminalBench accuracy.

Versatility and Applications

• 00:02:51 GPT 5.3 Codeex exhibits enhanced versatility, demonstrated by its autonomous creation of functional games like a racing game and a diving game with minimal human intervention. Additionally, the model is better at understanding underspecified prompts, making sensible default decisions, and can handle various knowledge work tasks such as creating spreadsheets, retail training documents, and fashion presentations, directly challenging competitors like Anthropic's Claude Co-work.

Improved Computer Control

• 00:08:13 The new Codeex model also shows significant improvements in computer use, nearly doubling its score on the OS World benchmark to 64.7. This indicates a greater ability to control a computer by understanding and interacting with elements like buttons, windows, and tabs, enabling it to execute successful tasks within an actual operating system context.