By ByteBot
On June 2, OpenAI announced that Codex — the company’s AI coding agent — now has 5 million weekly active users, up 6x since February. The headline buried inside that number: 20% of those users are not developers, and that segment is growing three times faster than the developer side. Alongside the usage figures, OpenAI launched Sites (AI-generated hosted web apps), Annotations (localized document edits), and six role-specific plugins connecting 62 enterprise applications. Codex is no longer being positioned as a developer productivity tool. It is becoming a workforce platform.
The Numbers That Matter
OpenAI’s accompanying report, “The Next Era of Knowledge Work,” published June 2, breaks down what non-developers are actually doing in Codex. Data analytics use is up 110% week-over-week. Research tasks are up 37%. Creating reports, spreadsheets, and knowledge artifacts is up 36%. These aren’t developers running benchmarks — these are analysts, marketers, and finance teams doing work that previously required an engineer in the loop.
That 3x growth rate for non-developers matters more than the absolute 20% figure. When a user segment is growing three times faster than your core segment, the product roadmap follows. That is not a criticism of OpenAI’s strategy — it is a statement about where resources will go and what features will get prioritized over the next 12 months. Developers using Codex should be aware of that dynamic.
What Actually Launched: Sites, Annotations, and Six Plugins
Three features shipped alongside the announcement. Sites lets Codex generate and deploy interactive web apps hosted by OpenAI infrastructure. The backend runs Cloudflare Worker-compatible ES modules with D1-style databases and R2 object storage — real persistent state, not static pages. The workflow: describe what you want, call @Sites in a thread, and Codex builds, tests, and deploys with a production URL. Available in preview for ChatGPT Business workspaces by default; Enterprise admins must enable it separately via role-based access control.
Annotations are more immediately useful for anyone editing documents in Codex. The feature lets you highlight a specific region — a block of cells in a spreadsheet, a slide element, a section of a document — and prompt Codex to change only that region. Before this, asking AI to update one chart in a financial model might regenerate the entire file, breaking formulas and custom formatting in the process. Annotations fix that by isolating execution within user-defined boundaries. The June 2 update expanded Annotations beyond code and Markdown to documents, spreadsheets, slides, and images.
The six role-specific plugins are the most explicit signal of OpenAI’s direction. Data Analytics connects Snowflake, Databricks, Hex, and Tableau. Creative Production connects Figma, Canva, Shutterstock, and Picsart. Sales connects Salesforce, HubSpot, Outreach, and Slack. Product Design, Public Equity Investing, and Investment Banking round out the set, together touching 62 applications and 110 specialized skills. More are coming: Corporate Finance, Private Equity, Marketing Strategy, and Legal are in development.
Related: Claude Code Dynamic Workflows: Parallel Agents, Real Cost Math
One Platform vs. Two Products
The strategic contrast with Anthropic is worth noting. Anthropic split its offering into two distinct products: Claude Code for developers, Claude Co-work for non-technical users. Two interfaces, two audiences, each optimized separately. OpenAI is betting on one interface for everyone — the same Codex thread that runs your async coding task also handles an analyst’s Snowflake query or a banker’s pitch deck. Neither approach is obviously correct.
On pure coding benchmarks, Claude Code still leads: Claude Opus 4.8 scores 88.6% on SWE-bench Verified. Codex with GPT-5.5 scores 82.7% on Terminal-Bench. Different benchmarks, different measurement methodologies — but the gap reflects a real difference in how each company is prioritizing developer-focused performance. If Codex’s roadmap is increasingly pulled toward knowledge worker use cases, that gap may widen. If OpenAI’s one-platform bet enables network effects between developer and non-developer usage, it may not matter.
What Developers Should Do Now
For developers evaluating Codex as a coding tool, nothing changed on June 2. As TechCrunch noted, Codex remains strong for async local repository workflows, and the Sites feature has genuine utility for rapidly deploying internal tools without a deployment pipeline. The question is whether OpenAI’s strategic pivot will affect future development priorities for developer-facing features. That is a watch-and-evaluate situation, not a reason to switch tools today.
For developers building tools and integrations, the plugin ecosystem is a real opportunity. Codex has 5 million weekly users. Getting a tool listed as a Codex integration exposes it to that audience, particularly within specific verticals. OpenAI is planning to expand the plugin set and the developer SDK for building custom plugins. If you build in finance, sales, analytics, or design, that is a distribution channel worth considering.
Key Takeaways
- OpenAI Codex reached 5 million weekly users on June 2; non-developers are 20% of that base and growing 3x faster than developers.
- Sites deploys AI-generated web apps with persistent state — useful for internal tools; available in ChatGPT Business preview now, opt-in for Enterprise.
- Annotations finally make AI document editing predictable: changes stay scoped to what you selected, leaving surrounding formulas and formatting intact.
- OpenAI’s one-platform strategy (developer and non-developer in one interface) differs from Anthropic’s two-product split — both are live bets on different architectures.
- The plugin ecosystem is a developer distribution opportunity: 62 apps today, with Corporate Finance, Legal, and others in development.
