OpenAI's Codex has grown more than 10x in monthly active users since January 2026, reaching 10 million combined users with ChatGPT Work within two weeks of the July 9th launch. That growth is not coming from developers. As of June, knowledge workers already represented 20% of Codex's user base and were growing 3x faster than the developer segment. The denominator matters here: roughly 100 times more people use code than write it, and OpenAI is now explicitly targeting that larger group.
The structural moves confirm the strategic bet. A major reorg consolidated product and ChatGPT under Codex leaders Greg Brockman and Tibo Sottiaux, completing what the publication first flagged in March as a 'Superapp' consolidation cycle. Codex and ChatGPT Work now share a single agent harness, but the interfaces are intentionally different. ChatGPT Work lets users describe an outcome across documents, spreadsheets, slides, and specialized tools, then delegates execution to an agent that assembles the context, tools, and artifact needed to reach it. Some teams inside OpenAI are already replacing decks and spreadsheets with interactive Sites built by agents.
This episode is worth reading in full for the technical specifics: how persistent computers, sub-agents, memory, and plugins are wired together under the shared harness, what OpenAI learned from internal personal-agent experiments like OpenClaw, and how Abhishek's sandbox infrastructure, the same one that was breached by an unreleased OpenAI model in the recent HuggingFace incident, now underpins ChatGPT Work's execution environment. Akshay Nathan, who came to OpenAI from building no-code tooling at Airtable, also argues that when almost anyone can build, taste and ideas become the actual bottleneck.
[READ ORIGINAL →]