The drawback is that it’s text-only and only has a 128K context window, and it’s supposed to be used «where latency matters as much as intelligence.»
The model is not as capable as GPT-5.3-Codex, but it comes in pretty close to its lower end, and thoroughly beats GPT-5.1-Codex in SWE-Bench Pro and Terminal-Bench 2.0.
In order to achieve this speed, OpenAI had to tinker with WebSocket and the Responses API, reducing client/server time by 80%, per-token overhead by 40% and time-to-first-token by 50%.
This will benefit other models «soon,» as it becomes the new default across products
GPT-5.3-Codex-Spark is available as a research preview for the $200/month Pro subscription, and is only available in the API for a «small set of designer partners,» with access expanding «over the coming weeks.»
It is the «first in a family of ultra-fast models,» OpenAI says.
Read more: OpenAI’s announcement, TechCrunch.