AI
OpenAI's Codex surges past 8 million users in five months after unified ChatGPT launch
The rapid adoption has forced OpenAI to scale infrastructure and adjust features, while intensifying competition with Anthropic's Claude and open-source alternatives.
The New Stack
OpenAI's Codex, now integrated into a unified ChatGPT desktop application following the launch of GPT-5.6 last week, has exploded from fewer than 1 million weekly active users in February to 8 million by early Sunday. The acceleration—reaching 5 million in early June, 6 million by mid-July, and doubling within days of the new release—represents growth few enterprise platforms ever experience.
The launch and infrastructure challenges#
The rollout combined three major changes: merging the standalone Codex app into ChatGPT's desktop client, launching ChatGPT Work as a new agentic interface for knowledge workers, and discontinuing the Atlas browser. The strategy backfired on capacity—traffic doubled OpenAI's previous peak within two days, exposing scaling bottlenecks.
Tibo Sottiaux, the engineering lead for Codex, detailed the team's response publicly. Key moves included:
- Optimizing inference pipelines to expand capacity by roughly 10% per user
- Reducing the context window from 372,000 to 272,000 tokens (the larger setting had created unexpected billing complications)
- Disabling experimental reasoning-effort configurations
- Reining in multi-agent behavior at higher reasoning tiers
- Temporarily lifting usage caps for paying subscribers
The context window adjustment sparked debate in the developer community—some viewed it as a hidden downgrade, while others appreciated the operational transparency. CEO Sam Altman framed the moves as a commitment to user respect, subtly positioning OpenAI against competitors he suggested treat customers poorly.
How the competition responded#
Anthropic moved quickly in response. Within hours of OpenAI's 7-million-user announcement, the company extended discounted pricing for Claude Fable 5 through mid-July and increased Claude Code's weekly usage allowance by half. Fable 5's performance has climbed to second place on Arena's agent leaderboard, with developers increasingly seeing OpenAI as reclaiming the coding AI lead.
Yet the picture remains complex. Products like Cognition's Devin Fusion, powered by Fable 5, can deliver lower per-task costs than even pricier models simply by reducing wasted effort—in 81% of Fable-driven runs, the primary model makes no code edits, avoiding redundant work downstream.
