OpenAI has introduced GPT-5.3-Codex-Spark, a new model designed for real-time coding. It’s smaller than its predecessor, GPT-5.3-Codex, and optimized for speed, delivering over 1000 tokens per second on low-latency hardware. This release marks a collaboration with Cerebras, focusing on improving responsiveness for developers. Codex-Spark allows for real-time collaboration, enabling users to make immediate adjustments and see results without delay. The model features a 128k context window and is currently text-only. During the research preview, it has its own usage limits, separate from standard models, to manage demand. Codex-Spark enhances the user experience by minimizing latency across the request-response process, achieving an 80% reduction in overhead for client-server interactions and cutting time-to-first-token by 50%. These improvements are expected to benefit all models in the future. Codex-Spark operates on Cerebras' Wafer Scale Engine 3, designed for high-speed inference. While GPUs remain essential for training and broader usage, this new infrastructure aims to tighten the feedback loop for coding tasks. The model is available today for ChatGPT Pro users and in API form for select partners, with plans for broader access as OpenAI refines its performance based on developer feedback. This launch sets the stage for future models that blend real-time collaboration with longer-term project execution.