OpenAI has made a bold move by sidestepping Nvidia with its innovative coding model, Codex-Spark, which runs on Cerebras' Wafer Scale Engine 3. This chip, the size of a dinner plate, has been the focus of Cerebras' business since at least 2022. But what's truly remarkable is the speed at which Codex-Spark operates. While 1,000 tokens per second might seem modest, it's actually impressive compared to Cerebras' own benchmarks. The company has measured 2,100 tokens per second on Llama 3.1 70B and reported 3,000 tokens per second on OpenAI's gpt-oss-120B model, suggesting that Codex-Spark's lower speed is due to the overhead of a larger or more complex model.
AI coding agents have had a breakthrough year, with tools like OpenAI's Codex and Anthropic's Claude Code reaching new levels of usefulness. OpenAI, Google, and Anthropic have been racing to ship more capable coding agents, and latency has become the key differentiator. A model that codes faster allows developers to iterate faster, which is crucial in the fast-paced world of software development.
OpenAI has been rapidly iterating on its Codex line, releasing GPT-5.2 in December after CEO Sam Altman's internal 'code red' memo about competitive pressure from Google. Just days ago, they shipped GPT-5.3-Codex. This rapid development cycle is a testament to OpenAI's commitment to staying ahead in the AI coding agent race.
But the story doesn't end there. OpenAI has been diversifying away from Nvidia, signing a massive multi-year deal with AMD in October 2025 and a $38 billion cloud computing agreement with Amazon in November. They've also been designing their own custom AI chip for eventual fabrication by TSMC. This move away from Nvidia is significant, especially given the planned $100 billion infrastructure deal with Nvidia that fizzled out.
Despite the speed of Codex-Spark, it's important to note that speed may come at the cost of accuracy. For developers who spend their days inside a code editor waiting for AI suggestions, 1,000 tokens per second might feel like running a rip saw. But the key is to watch what you're cutting, and with the right balance of speed and accuracy, Codex-Spark could be a game-changer for rapid prototyping and code development.
