Executive Summary
OpenAI has launched GPT-5.3-Codex-Spark, a new, smaller model designed for ultra-low-latency, real-time coding tasks. Developed in partnership with Cerebras and running on their specialized hardware, Codex-Spark aims to complement existing models by enabling rapid, interactive iteration with near-instant responses of over 1000 tokens per second. The model is being rolled out as a research preview to ChatGPT Pro users within the Codex environment to gather feedback on this new interactive paradigm.
Key Takeaways
* Product: GPT-5.3-Codex-Spark, a smaller, specialized version of the GPT-5.3-Codex model.
* Primary Function: Optimized for speed and real-time collaboration, allowing developers to interrupt, redirect, and iterate with the model with near-instantaneous feedback.
* Performance: Delivers over 1000 tokens per second, powered by Cerebras' Wafer Scale Engine 3 AI accelerator for ultra-low latency inference.
* Initial Specs: Launches with a 128k context window and is text-only.
* Availability: Rolling out as a research preview to ChatGPT Pro users via the latest Codex app, CLI, and VS Code extension, with separate rate limits.
* Platform-wide Improvements: Latency enhancements, including a new persistent WebSocket connection, have been implemented that benefit all models, reducing roundtrip overhead by 80% and time-to-first-token by 50%.
Strategic Importance
This launch signifies OpenAI's strategy to bifurcate its model offerings, creating specialized tools for different workflows. By addressing the critical bottleneck of latency, OpenAI aims to make AI collaboration more natural and interactive, complementing its powerful but slower models designed for complex, long-running tasks.