OpenAI sidesteps Nvidia with unusually fast coding model on plate-sized chips (arstechnica.com)

OpenAI has announced the release of its first production AI model, GPT-5.3-Codex-Spark, designed to operate on non-Nvidia hardware, specifically leveraging Cerebras chips. This new coding model astonishingly delivers code at a rate exceeding 1,000 tokens per second—approximately 15 times faster than its predecessor, GPT-5.1-Codex-mini. In contrast, Anthropic’s Claude Opus 4.6 in its premium mode achieves about 172 tokens per second. OpenAI's head of compute, Sachin Katti, emphasized the importance of this partnership with Cerebras and the model's rapid inference capabilities. Significantly, this release marks a pivotal shift in OpenAI's hardware strategy, moving away from Nvidia dominance in AI infrastructure. Codex-Spark, focused on speed rather than depth of knowledge, is available for ChatGPT Pro subscribers and functions as a text-only model specialized for coding tasks. Early performance metrics suggest that Spark excels in competitive benchmarks against older models, though independent verification of these results has not been provided. By offering unprecedented coding speeds, OpenAI positions itself strongly in the ongoing race to enhance AI coding agents, aiming to meet growing demands in software development efficiency.