OpenAI retires its high-speed model GPT-5.3-Codex-Spark after 7 months

OpenAI has retired its fastest model, GPT-5.3-Codex-Spark, which was launched on February 12, 2026. The model, designed for real-time programming, was the first to operate outside NVIDIA's technology stack, using Cerebras' WSE-3 chip. It was capable of processing 1200 tokens per second, but its performance in benchmarks was significantly lower than the full version of GPT-5.3-Codex, scoring 58.4% on Terminal-Bench 2.0 compared to 77.3% for the full model. Tibo, head of Codex, cited declining usage and the availability of better models as reasons for the retirement. Cerebras' Ultrafast mode, introduced on August 13, 2026, allowed the flagship model GPT-5.6 Sol to run directly on the wafer, achieving a 5.6 times speedup. This marked a shift in OpenAI's strategy, moving away from distilled models toward full-scale performance. Over the past three months, OpenAI has retired several older models, including GPT-5.2, GPT-5.3-Codex, GPT-5.4, and GPT-5.4 Mini, with users transitioning to the 5.6 generation. The company is now adopting a new naming convention with models like Sol, Terra, Luna, and Astra.