OpenAI Codex has reached 7 million active users, adding roughly one million in a single day. The jump came ten days after the launch of GPT-5.6 Sol, which brought the frontier model directly into the coding agent.

Tibo, Codex’s product lead, announced 6 million users on July 12. By July 13, 7 million. Back in January, Codex was between 550k and 700k users: ten-fold growth in six months.

The metric that matters is trajectory. When a tool gains a million users per day, it signals that people writing code for work are integrating it into their routine, not trying it once and moving on. Coding agents are transitioning from proof-of-concept to daily use.

For anyone evaluating agentic tools, mass adoption is a maturity signal. Codex with GPT-5.6 Sol has reached the point where real-world feedback outweighs internal testing, and where limits surface in production (as we saw with the bug that deleted the home directory).

Another striking data point: the last public number for Claude Code is 2 million users, from February. Since then, silence. Anthropic has shifted some coding traffic to Claude Tag, a Slack bot, making comparison difficult. On visible metrics, Codex is ahead of Claude Code.

In detail

To understand the trajectory, look at the starting numbers. In January 2026, OpenAI reported roughly 550-700k active Codex users. By March, Fidji Simo mentioned 2 million. By July, 7 million. Adoption accelerated with GPT-5.6’s launch on July 9: in 48 hours (July 10-12), Codex went from 5 to 6 million users, then hit 7 million the next day.

Comparison with Claude Code. Anthropic’s last public data is from February: 2 million weekly users, with a doubling in the preceding six weeks. Since then, no numbers. Latent Space notes that Anthropic may have shifted some coding traffic to Claude Tag, an assistant integrated into Slack, which has usage metrics that don’t easily compare to command-line tools. Silence can mean many things, but the fact remains: on public numbers, Codex is ahead.

What’s driving growth. GPT-5.6 Sol is the model OpenAI built with agents in mind: multi-step planning, tool use, sub-agent delegation. JetBrains made Codex its recommended agent. OpenAI temporarily lifted hourly usage limits for Plus, Business, and Pro plans—a sign demand exceeded expectations. Sam Altman spoke of 2.5x growth in a week across Codex and ChatGPT Work.

One insight from sources: model quality matters less and less as a differentiator. What matters more is the quality of the harness—the architecture surrounding the model, orchestration, tools, memory, and controls. As more users enter, orchestration reliability and visibility into agent behavior become the real winners (LangChain added tracing for Codex, Cursor, Copilot, and others).

Limits of what we know. The numbers come from OpenAI itself, not independent measurement. “Active users” includes both one-time testers and those using it eight hours a day, and OpenAI hasn’t published the distribution. Growth was also boosted by removing usage limits and perks like the “banked reset” distributed to everyone to celebrate 7 million, which can temporarily inflate metrics.

Security questions remain open. Growing this fast means more surface for errors. The bug that deleted the home directory when sandboxing was removed for speed is an example: more users entering means more edge cases surface in production rather than internal tests.

To follow the thread: on July 18 we covered the Codex and GPT-5.6 integration with 2.5x weekly growth and JetBrains as the recommended agent. The trajectory has only accelerated since.