OpenAI has integrated Codex’s agent capabilities into the newly launched ChatGPT Work. Less than two weeks after the July 9, 2026 launch, the company reported that ChatGPT Work and Codex together had reached about 10 million users. Internal usage patterns at OpenAI — and a >10x increase in Codex monthly active users since January 2026 — showed that coding agents were rapidly extending into knowledge work.
What happened and why it matters
- Codex, originally framed as a developer tool, now runs on the same agent harness as ChatGPT Work. That means all ChatGPT Work users are effectively using the Codex agent infrastructure.
- OpenAI observed unexpectedly strong adoption of Codex among non-developers inside the company: teams in strategic finance, marketing and other knowledge functions began using Codex and described it as giving them a kind of “superpower.”
- In June 2026 OpenAI said knowledge workers already made up roughly 20% of Codex’s users and that this segment was growing more than three times faster than developers.
Product and organizational decisions
Those internal patterns, together with a broader “superapp” consolidation, led OpenAI to merge experiences: the underlying agent harness is shared, but user-facing differences remain (UX choices, Git visibility, sandboxing defaults). OpenAI’s product stance is to provide a strong opinionated default that works for most users while enabling power users to opt into deeper reasoning, Ultra or multi-agent modes when necessary.
How the product adapts to knowledge work
- Agents and sub-agents: ChatGPT Work supports breaking complex tasks into parallel sub-agents; design trade-offs involve how much of those internals to display and how to balance cost, latency and transparency.
- Artifacts and Sites: instead of relying only on documents, spreadsheets and decks, ChatGPT Work lets users produce higher-fidelity artifacts and hosted Sites that teams can iterate on collaboratively.
- Memory and Chronicle: ChatGPT’s Memory (Memory V3) and Chronicle features let the system accumulate longer-term, personal and project context; OpenAI emphasizes these as tools for better contextual retrieval rather than replacements for human judgment.
Models, defaults and practical guidance
OpenAI aims to make the default configuration the best option for most users. Multiple model classes and a reasoning slider exist to let users trade speed for thoroughness; advanced modes and model choices are available to power users but are not necessary for typical tasks. The team’s advice: start with the default and adjust only if you need higher reasoning or more deterministic behavior.
Power-user advice from OpenAI’s product lead
Akshay Nathan (Head of Core Product Engineering / Productivity at OpenAI) recommends expanding what you imagine the tools can do, storing artefacts and context in the product’s library, and iterating — because the more domain context the agent has, the more proactively useful it becomes. He also urges retrying tasks models could not handle months earlier, since model capabilities improve rapidly.
Safety, trust and human oversight
OpenAI stresses that in sensitive domains (e.g., performance reviews) agents are best used to gather context, not to replace managerial judgment. Permissions, sandboxing and careful defaults are critical because ChatGPT Work agents can access deeply personal and internal company information via plugins and local file access.
OpenClaw, persistent environments and personal agents
Personal-agent projects like OpenClaw influenced ChatGPT Work. ChatGPT Work provides a persistent “computer” environment with file storage, scheduled tasks and automations; some internal users migrated personal use cases (meal planning, workouts, household management) from prototypes like OpenClaw into ChatGPT Work’s environment.
Measuring productivity: motion versus progress
Nathan argues that old proxies (commits, tokens, PRs) are less informative now. Instead, teams should measure whether they are getting to outcomes — “quality at-bats” that cycle from idea to validation. A key trap is conflating motion (lots of activity) with meaningful progress toward goals.
Scale and market opportunity
OpenAI’s estimates point to a much larger market in knowledge work than in development: roughly O(5 billion) potential knowledge workers versus O(50 million) developers. The company’s product roadmap sequences from developer-first tools to general knowledge work and then toward broader consumer use.
Key dates and numbers
- July 9, 2026: ChatGPT Work announced; within two weeks OpenAI reported ~10 million combined users for ChatGPT Work and Codex.
- Jan → mid‑2026: Codex MAU increased by more than 10×.
- June 2026: knowledge workers were about 20% of Codex users and growing >3× faster than developers.
Bottom line
Codex’s unexpected adoption beyond engineering prompted OpenAI to converge its agent experiences into ChatGPT Work on a shared agent harness. The company has chosen to keep opinionated UX and safety defaults for different contexts while enabling advanced configurations for power users. OpenAI’s immediate challenge is getting the default experience to unlock agentic knowledge work for many more people while preserving oversight, privacy and predictable behavior as usage scales.
