OpenAI has rolled out its GPT-5.3-Codex-Spark on Cerebras hardware, making the first production AI model from the company to run on non-Nvidia chips. Spark targets coding tasks and is capable of delivering about 1,000 tokens per second, a pace OpenAI says is roughly 15 times faster than its predecessor. The model ships with a 128,000-token context window and, at launch, handles text-only output.
“Cerebras has been a great engineering partner, and we’re excited about adding fast inference as a new platform capability,” said Sachin Katti, OpenAI’s head of compute.
Codex-Spark is offered as a research preview to ChatGPT Pro subscribers ($200/month) via the Codex app, a command-line interface, and a VS Code extension, with API access rolling out to select design partners. Spark is positioned for speed, not the broader, agentic coding tasks of the full GPT-5.3-Codex model, and is designed specifically for coding use cases.
OpenAI describes Spark as a text-only model tuned for speed, while the full GPT-5.3-Codex aims to handle heavier, more complex coding tasks. Spark’s 128k-token context window makes it well-suited for rapid prototyping and iterative coding workflows, though it may sacrifice some depth of knowledge compared with the larger model.
On SWE-Bench Pro and Terminal-Bench 2.0, Spark reportedly outpaces the older GPT-5.1-Codex-mini by completing tasks in a fraction of the time. OpenAI has not provided independent validation of these numbers. By contrast, independent benchmarks on Nvidia hardware show other OpenAI models topping out well below Spark’s reported speed, with tokens-per-second in the low hundreds for models such as GPT-4o and its variants.
