Cerebras Systems CEO Andrew Feldman has confirmed that OpenAI’s GPT-5.4 model is now running on Cerebras hardware, with plans to bring GPT-5.5 onto the platform in the near term. The announcement deepens what has already become one of the most consequential infrastructure partnerships in AI.
The OpenAI-Cerebras relationship was formalized on January 14, 2026, when the two companies announced a multibillion-dollar agreement reportedly exceeding $10 billion. The deal covers the provision of up to 750 MW of compute capacity over a three-year window stretching through 2028 or 2029, focused on ultra-low-latency inference for OpenAI’s flagship products, including ChatGPT.
The first major milestone came on February 12, 2026, when the GPT-5.3-Codex-Spark model launched on Cerebras hardware. That model clocked in at over 1,000 tokens per second. Now, with GPT-5.4 confirmed live and GPT-5.5 in the pipeline, Cerebras is becoming the preferred inference engine for OpenAI’s entire next-generation lineup.
Every single time someone asks ChatGPT a question, that’s an inference call. Every API request from a developer building on OpenAI’s models triggers inference. The over-1,000-tokens-per-second benchmark set by GPT-5.3-Codex-Spark on Cerebras hardware signals that real-time AI interactions are becoming table stakes for applications like live coding assistants, real-time translation, and conversational agents.
