Codex hit 10 million users in two weeks, and 20% of them aren't developers

On Latent Space July 28, OpenAI core product engineering lead Akshay Nathan said Codex and ChatGPT Work combined reached 10 million users within two weeks of the July 9 launch, with monthly actives up more than 10x since January 2026.

The number that should reframe your product thinking: knowledge workers are already roughly 20% of Codex's user base and growing 3x faster than developers.

Architecturally, both products run the same underlying agent harness behind different interfaces. Codex leans on Git integration and diffs. Work optimizes for non-technical discovery and inherits memory from the main ChatGPT product for persistent cross-session context. One harness, two surfaces, and the non-developer surface is the one compounding.

Once you notice this, today's other releases stop looking like a coincidence. Gemini Spark launched in India July 29 as a persistent personal agent running on Google's cloud rather than the device, so it keeps working with your laptop shut. Gemini 3.6 Flash drives it, it connects natively to Gmail, Docs, and Sheets with zero setup, and it handles standing tasks: triaging email into documents, conditional calendar events, subscription monitoring, drafting emails for review. Google gates the high-stakes actions, saying Spark asks before spending money or sending email. Perplexity's Personal Computer for Windows landed July 27 for Max subscribers, orchestrating 15+ models across local Word and Excel files, Microsoft 365, and the web. Their lead demo is pointing it at a to-do list in Notepad. Not a repo. Notepad.

On the commercial side, Viktor tops today's Product Hunt at 126 upvotes selling "an AI coworker that actually does the work," living in Slack or Teams as an @-mentionable teammate wired to 3,000+ tools. It hit a €12.9M run rate within 10 weeks of launching in February, 2,000+ organizations, $75M Series A led by Accel. The unit being sold is a coworker, which prices against salary rather than against a per-seat SaaS line. And Andrew Ng's OpenWorker hit 10,380 stars and 1,349 forks in nine days as an open desktop coworker that produces finished deliverables instead of chat, with every write, send, or shell command approval-gated and unattended runs parking their asks in an inbox.

I've been building agent tooling with an implicit assumption that the user reads a terminal. The 3x growth differential says that assumption has an expiration date. If your agent product's surface area is a CLI, the harness underneath is probably fine but the interface is addressing the slower-growing half of the market.