Codex-Spark: OpenAI's Fast Coding Model
OpenAI is releasing GPT-5.3-Codex-Spark, a fast but imprecise coding model. It runs on a dedicated Cerebras chip.
(Image: Peshkova / Shutterstock.com)
A first model for real-time coding, that's how OpenAI describes the newly released GPT-5.3-Codex-Spark. It is a research preview and uses a Cerebras chip for the first time.
Codex-Spark is said to be particularly fast – specifically capable of delivering 1000 tokens per second. However, OpenAI also writes in a blog post that this can come at the expense of quality – at least that's what the Terminal-Bench 2.0, which aims for accuracy, shows. Nevertheless, thanks to its speed, new and different interactive work with the model should be possible. For example, Codex-Spark can also be interrupted or redirected in real-time, it is said. However, there is no automatic preview, for instance. Fundamentally, only text is processed, and the model has a 128K context window.
In January, OpenAI announced its partnership with the Californian chip designer Cerebras. They have since then been working intensively on a chip designed for inference, meaning it can execute AI algorithms particularly quickly. Until now, OpenAI had relied on AI accelerators from Nvidia. However, Codex-Spark doesn't sound fully mature yet. The blog post states that they want to release the model for early experiments while working on the end-user experience, among other things. Initially, there are also special rate limits, including that usage can be fundamentally restricted with many accesses. ChatGPT Pro users with the Codex app, the CLI, and the VS Code extension have access.
Videos by heise
OpenAI also announces that Codex-Spark is intended to be the first model in a new “ultra-fast model family.” Multimodality and further capabilities are expected to follow.
Just a few days ago, OpenAI released GPT-Codex-5.3, a new model with coding capabilities. This is also said to be faster than its predecessor. However, this is not about real-time processing but about minute-long thinking processes to complete tasks. Additionally, there is the Codex app, a command center for AI workflows.
Sam Altman jokes on X that the new, fast model would bring him joy. He refers to the 2010s TV show where Marie Kondo helped people declutter with the English sentence “It sparks joy for me.” The tidying and minimalism expert asked each item if it brought the owner joy – only then could it be kept. However, Kondo has long since abandoned her own “spark” style and, according to her statements, prefers to allow a little chaos.
(emw)