Ever since openai switched to websockets in the codex harness the caching has been crazy good. I've kept track of how much of my input has been cached. Before openai websockets i was sitting around 25-40% of input tokens being cached ever since the websocket update my caching in harnesses making use of it has increased to around 85%, and is around 95%-98% in Codex. Claude code in comparison gets about 45% caching over long-running tasks; their cache busts a lot more often.
This is most likely why even though gpt-5.5 and gpt-5.4 were so much more expensive than gpt-5.3-codex it hasnt felt that much more expensive if you compare it to what caching used to be like about 3 months ago. Another very cool thing openai does that anthropic doesn't. Cache writes for openai is free vs anthropic who bills it at a higher price than regular token ins lol
Openai actually been cooking man!