OpenAI Codex-Spark: Revolutionizing Coding with Cerebras Hardware
In a groundbreaking move, OpenAI has unveiled GPT-5.3-Codex-Spark, a cutting-edge AI model designed to revolutionize coding experiences. This model is a significant departure from traditional hardware, as it runs on Cerebras wafer-scale chips, offering unprecedented speed and efficiency.
The new model boasts an impressive 1,000 tokens per second, a 15x improvement over previous versions. This breakthrough in speed is a game-changer for real-time coding assistance, allowing developers to iterate and refine code with lightning-fast responses. OpenAI's focus on low latency and interactive coding workflows ensures that Codex-Spark provides an unparalleled coding experience.
But the benefits don't stop there. Despite its speed, Codex-Spark retains the ability to handle long-running processes, operating for extended periods without intervention. This dual capability of speed and endurance makes it a versatile tool for software engineers.
The model's performance was evaluated using SWE-Bench Pro and Terminal-Bench 2.0, benchmarks specifically tailored for software engineering tasks. Results showed that Codex-Spark outperformed its predecessors, GPT-5.1-Codex-mini and GPT-5.3-Codex, in a fraction of the time. OpenAI's commitment to reducing latency across the entire request-response pipeline will enhance the performance of all their models.
Under the hood, OpenAI has made significant optimizations. They streamlined the response streaming process, rewrote critical components of their inference stack, and improved session initialization. These changes have led to a remarkable 80% reduction in per-client/server roundtrip overhead, a 30% decrease in per-token processing time, and a 50% reduction in time-to-first-token. These improvements are set to become the default for all models.
Codex-Spark runs on Cerebras' Wafer Scale Engine 3 accelerators, which are designed for low-latency, high-speed inference. However, OpenAI emphasizes that this doesn't signal a departure from GPUs. Instead, Cerebras accelerators can be combined with GPUs to leverage the strengths of both architectures.
The announcement has sparked online discussions, with some users prioritizing intelligence and reliability over speed. Tystros, for instance, expressed a willingness to wait for better results, even if it takes longer. Others, like stobak, highlighted the potential cumulative cost of faster models, which may incur repeated iterations. Nicholas Van Landschoot, on X.com, questioned the dramatic speed claims, measuring improvements closer to 1.37x rather than 15x in practical benchmarks.
Codex-Spark offers a 128k context window and text-only support, with plans to introduce faster models with larger contexts based on community feedback. OpenAI's commitment to continuous improvement and collaboration with Cerebras ensures that the future of coding assistance is bright and innovative.
