GPT-5.3 Codex Spark

Introducing GPT-5.3-Codex-Spark—our first real-time coding model. 15x faster generation, 128k context, now in research preview for ChatGPT Pro users.

OpenAI released GPT-5.3-Codex-Spark as a research preview, describing it as a smaller variant of GPT-5.3-Codex and its first model purpose-built for real-time coding inside Codex. The launch marks the first milestone in OpenAI's January-announced partnership with Cerebras, and the model is served on Cerebras hardware t At launch, Codex-Spark is text-only with a 128k context window and has its own rate limits that do not count against standard Codex limits, though OpenAI warned of temporary queuing under high demand. It is tuned for interactive use, making minimal targeted edits by default and skipping automatic test runs unless asked

OpenAI is retiring GPT-5.3-Codex-Spark after roughly seven months of availability, according to a September 13, 2026 report attributing the announcement to Thibault Sottiaux, head of Codex, who confirmed the decision on X. Sottiaux cited steadily declining usage and noted that OpenAI already fields clearly superior mod The model, launched February 12, 2026, was OpenAI's first production inference deployment outside Nvidia GPUs, running on Cerebras WSE-3 wafer-scale chips and delivering more than 1,000 tokens per second for real-time coding. It scored 58.4% on Terminal-Bench 2.0 versus 77.3% for the full GPT-5.3-Codex, reflecting the

In a major shift in its hardware strategy, OpenAI launched GPT-5.3-Codex-Spark, its first production AI model deployed on Cerebras wafer-scale chips rather than traditional Nvidia GPUs. The new model

OpenAI released the Codex-Spark coding model to help developers to make targeted edits, adjust logic, refine interfaces, and more.

An independent engineering comparison published February 14, 2026 and updated June 4, 2026 characterizes GPT-5.3-Codex-Spark as a distilled, Cerebras-accelerated variant of OpenAI's Codex 5.3, purpose-built for low-latency code generation at 1,000+ tokens per second with a 128k context window, available as a research p The review supplies one of the few concrete tradeoffs available: on SWE-Bench Pro Spark scored 56% versus Codex 5.3's 72%, and a hands-on snake-game build took Spark about 50 seconds with a left-wall collision blind spot and a memory-leaking reset, while Codex 5.3 took roughly six minutes but handled edge cases correct

Cerebras confirmed on its company blog that OpenAI's GPT-5.3-Codex-Spark launched in research preview on February 12, 2026 running on the Cerebras Wafer-Scale Engine at over 1,000 tokens per second, marking the first release in the Cerebras–OpenAI collaboration. The post frames Spark as the response to the shift toward Cerebras reports that on agentic software engineering benchmarks SWE-Bench Pro and Terminal-Bench 2.0, Codex-Spark produces more capable responses than GPT-5.1-Codex-mini while completing tasks in a fraction of the time, and includes a quote from OpenAI's Sachin Katti calling wafer-scale compute a new platform capabili