I didn't even have time to post about yesterday's GPT-5.3 and the Codex for Windows release before they rolled out another major update. The new GPT-5.4 shows a serious leap in coding, web search, and other agentic tasks.
Benchmark numbers:
— SWE-Bench Pro — 57.7%
— OSWorld-Verified — 75.0%
— BrowseComp — 82.7%
The main new feature: you can now interrupt generation mid-response to add instructions or redirect the model's thinking on the fly (something already available in e, for example).
Access to GPT-5.4 opens this week on web and Android (with the Pro version right away). iOS will come a bit later.