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 offers delivers improved throughput and low-latency, enabling a real-time, interactive coding experience, says the company.
We’re sharing Codex-Spark on Cerebras as a research preview to ChatGPT Pro users so that developers can start experimenting early while we work with Cerebras to ramp up datacenter capacity, harden the end-to-end user experience, and deploy our larger frontier models.
Codex-Spark runs at roughly 1,000 tokens per second, about 15× faster than earlier versions, making live coding assistance and rapid iteration much more responsive. OpenAI says the new model was designed “specifically for working with Codex in real-time—making targeted edits, reshaping logic, or refining interfaces and seeing results immediately”.
To enable real-time coding, OpenAI optimized Codex-Spark for low latency and interactive coding workflows rather than deep reasoning or general-purpose tasks. Despite this focus on speed, the model retains its predecessor’s ability to handle long-running processes, operating for “hours, days, and weeks without intervention”.
OpenAI says that GPT‑5.3‑Codex‑Spark demonstrated its performance on SWE-Bench Pro and Terminal-Bench 2.0, two benchmarks tailored for software engineering tasks, achieving results between GPT-5.1-Codex-mini and GPT-5.3-Codex but in a fraction of the time. The company also notes that end-to-end improvements implemented to reduce latency across the full request-response pipeline will benefit all their models.
Under the hood, we streamlined how responses stream from client to server and back, rewrote key pieces of our inference stack, and reworked how sessions are initialized so that the first visible token appears sooner and Codex stays responsive as you iterate.
Among other enhancements, OpenAI introduced a persistent WebSocket connection and several optimizations in the Responses API. Overall, these improvements reduced per client/server roundtrip overhead by 80%, per-token processing time by 30%, and time-to-first-token by 50%. These changes will become the default for all models, OpenAI says.
Codex-Spark runs on Cerebras’ Wafer Scale Engine 3 accelerators, which are particularly suited to low-latency, high-speed inference. However, this does not signal a departure from GPUs as the core of their training and inference pipeline, according to OpenAI. Cerebras accelerators can also be combined with GPUs to achieve the best of both architectures.
OpenAI’s announcement sparked significant online discussion. Some reddit users emphasized a preference for “maximum intelligence and reliability” over speed, with Tystros commenting: “[if the results are better when it takes one hour to complete a task, I happily wait one hour]”(https://www.reddit.com/r/codex/comments/1r30pvl/comment/o50tpmu/). User stobak highlighted that it is easy to underestimate the cumulative cost of repeated iterations that faster models can incur.
Nicholas Van Landschoot observed on X.com that speed improvements are not as dramatic as claimed, measuring closer to 1.37x rather than 15x in practical benchmarks. He explains that the 15x figure comes from comparing Codex-Spark to a specific configuration of Codex, x-high, which is used to force longer reasoning time to increase accuracy.
Codex-Spark provides a 128k context window and text-only support, with plans to introduce faster models featuring larger contexts based on usage insights gathered from the developer community.
