OpenAI Introduced Cerebras-Powered GPT-5.3-Codex-Spark
New Model Focused on Fast Inference
OpenAI announced GPT-5.3-Codex-Spark, a lighter and faster version of the agentic coding model GPT-5.3-Codex. The company stated that it optimized this new model for development processes requiring fast iteration and real-time collaboration.
Cerebras Wafer Scale Engine 3 Integration
In the model's infrastructure, Wafer Scale Engine 3 (WSE-3) chips from OpenAI's hardware partner Cerebras are used. Equipped with 4 trillion transistors, these chips were designed to support workflows requiring extremely low latency. This move stands out as a concrete example of OpenAI's strategy of working with manufacturers such as Cerebras, AMD, and Broadcom by diversifying its Nvidia-dominated infrastructure.
Technical Specifications and Performance
Although GPT-5.3-Codex-Spark performs worse than the main model GPT-5.3-Codex in complex software engineering tasks, it offers an advantage in speed. The model, which ranked lower in SWE-Bench Pro and Terminal-Bench 2.0 tests, aims to preserve developers' creative flow by delivering more than 1000 tokens per second.
- Context Window: 128 thousand tokens.
- Data Type: Text-only support.
- Access: Research preview in the Codex app for ChatGPT Pro users and limited API access.
Cerebras CTO Sean Lie emphasizes that this model is a starting point for exploring new interaction patterns enabled by fast inference.