GPT-5.3-Codex-Spark Live!

[zemaj]

Wow it's fast! And very accurate for a fast model. I'll probably make this the default auto review/resolve model once widely available.

Full websocket support has been added to, so getting absolutely every bit of speed out of it!

[zemaj]

An excellent use case for it is letting it loop on errors. For example, if you have a program which is executing and showing errors, letting it loop on testing, patching, testing until resolved is very effective. So sticking with GPT-5.3-Codex High for general planning/execution and GPT-5.3-Codex-Spark for debugging.

[cbusillo]

I swapped to spark for auto-review. Do you suggest medium?

Would it make any sense to make spark frontline? It's currently not going to be picked for actual coding work as, is that your intent? I don't believe its exposed in the agent config.

[zemaj]

I use high for spark, but I don't think there's a huge difference between reasoning levels.

I'm experimenting at the moment with automatically switching between spark and codex in auto drive based on the task in that turn. It is currently in the agent list.

I do find that spark struggles with complex tasks (of course), but it's far better than flash models from google at the moment. It's more like low/medium -codex level at 15x speed which is just amazing. It opens up a lot of "behind the scenes" analysis at that speed.