OpenAI unveils GPT-5.3 Codex Spark, a coding model built for speed

On February 12, OpenAI introduced the GPT-5.3 Codex Spark, a revolutionary AI model engineered for rapid software development. This launch signifies a pivotal moment in the competitive landscape of AI tools, as it aims to facilitate instantaneous and interactive coding processes. The new model is a more compact iteration of the existing GPT-5.3 Codex system, and it is currently available as a research preview for ChatGPT Pro users and select collaborators. Optimized for performance, Codex Spark is designed to deliver responses in real-time when utilized on specialized low-latency hardware, boasting the capability to generate over 1,000 tokens per second. This launch also marks a significant milestone in OpenAI's collaboration with Cerebras, a chipmaker that focuses on high-speed computing solutions. Unlike larger models intended for extensive autonomous tasks, Codex Spark is finely tuned for swift code editing, logical refinement, and immediate user interaction. OpenAI emphasized that this model is crafted specifically for real-time usage with Codex, enabling users to make precise adjustments and observe results instantly. The system caters to both quick tasks and complex projects, allowing developers to adjust outputs dynamically as they are generated. With a context window of 128,000 tokens, the model currently operates in a text-only format. Although it is smaller than some cutting-edge models, OpenAI asserts that it excels in software engineering benchmarks like SWE-Bench Pro and Terminal-Bench 2.0, completing assignments significantly faster. This release aligns with a growing trend in the industry towards specialized models that emphasize responsiveness over sheer reasoning capability, especially in developer tools where latency can impact productivity. OpenAI has also revamped its infrastructure to minimize delays throughout the request process, implementing persistent WebSocket connections and enhancements to its Responses API, resulting in an 80% reduction in client-server roundtrip times and halving the time to receive the first token. Codex Spark is powered by Cerebras’ Wafer Scale Engine 3, a dedicated AI accelerator that is optimized for ultra-fast inference. This hardware setup complements traditional GPU infrastructures by focusing on minimal latency. Sean Lie, co-founder and CTO of Cerebras, expressed enthusiasm about the collaboration with OpenAI and the developer community, stating that this preview represents just the start of exploring new interaction paradigms and use cases enabled by fast inference technology.