In July 2026, OpenAI’s coding agent Codex together with ChatGPT Work crossed 10 million weekly active users. Six months earlier, that number was under 5% of competitor Claude Code’s.
Altman recently called the catch-up campaign a “kamikaze mission” — and the point of kamikaze was never courage; it was willingness to pay any price. To take back the coding-agent track, OpenAI shut down Sora, killed the Atlas browser, raised internal coding compute 100-fold, and tolerated a subsidy system with almost no ceiling.
The result stunned the industry. But what deserves analysis is not the growth rate — it is which deep rules of AI product competition the sprint reveals.
Why Altman Valued This Fight Above Sora
Start with why losing was unacceptable. In the interview, Altman revealed a key internal metric: the RSI (Recursive Self-Improvement) index — a measure of how well models participate in their own iteration, spanning troubleshooting, low-level optimization, pretraining improvements, and post-training process refinement. From GPT-5.5 to 5.6, the index jumped from 41.7 to 57.9. A coding agent is not merely a product sold to customers; it is the tool that manufactures the next generation of products.
That explains the Sora sacrifice. When a product has the infrastructure property of “intelligence producing intelligence,” its strategic value stops being revenue and becomes the variable that decides generational gaps between models. Sora, however successful, is an application in video generation. A coding agent directly participates in building the next model.
It also exposes a bottom-up change in the industry: the boundary between product and production tool is dissolving. When your user-facing product (Codex) doubles as the infrastructure for your next product (Codex improving GPT), it naturally outranks every pure application.
How Codex Actually Won
The answer: not quality — a combination of speed, distribution, and subsidies.
Speed. On Artificial Analysis’s coding-agent evaluation, Claude Code averages 23.6 minutes per task; Codex needs 10.2. In blind tests Claude Code still leads on code quality with a 67% win rate — but developers vote with their feet, and the logic is blunt: most daily work is adding tests, adjusting interfaces, fixing edge cases, where finishing faster beats finishing better.
Underneath the speed gap is token efficiency. On the same task, Claude Code consumes an average of 21.8 million tokens across 153 model interactions; Codex uses 13.2 million across 114. OpenAI’s explanation: GPT-5.6 was trained to take shortcuts — in tool-heavy tasks it prefers writing a snippet of code to handle intermediate results rather than re-feeding every tool return back to the model for judgment.
Distribution. Codex ships inside existing ChatGPT plans — free users get trial quotas, paid users get bigger ones. Trying Codex requires no new registration, no new account, no new pricing model to learn. The friction is nearly zero.
Subsidies. Codex’s product lead Tibo reset usage quotas for all users roughly every 9 days — tightening to every 3 days in July — so often that someone built codex-resets.com just to track the resets. More extreme: if your quota runs out mid-task, Codex finishes the job before stopping. Subscription expired? Pro privileges quietly persist for days.
The combo shows up internally too: by end of June 2026, 98% of OpenAI employees were working through Codex — well beyond engineering, with over 85% using it cross-functionally weekly. Research-scenario usage grew 56× in seven months, and requests for 8-hour autonomous agent runs grew 10×.
The Price of the Sprint
The catch-up was expensive, and not only in a shut-down Sora and a dead Atlas browser.
GPT’s writing regressed. Back in January, someone asked Altman why GPT-5 codes brilliantly but writes worse. He didn’t dodge: most of the effort went to reasoning, coding, and engineering; writing and conversation waited. A model cannot sprint in every direction at once.
Compute spending is startling. The $200/month plan frequently delivers well over $1,000 of model usage. The subsidy generosity borders on absurd — quotas reset weekly, reset on bugs, reset on version updates.
The business still isn’t profitable. Claude Code passed $2.5 billion in annualized revenue by February 2026. Codex crossed $1 billion only at the end of January. Market share caught up; revenue and unit economics remain unproven.
The Structural Rules Beneath the Story
-
When the product is infrastructure, resource concentration overrides everything. An early misjudgment on code completion cost OpenAI its entire first-mover advantage. Once the strategic meaning of RSI became clear, Altman cut every other good thing without hesitation.
-
In AI tool markets, speed wins daily usage over quality. Blind tests still favor Claude Code on quality, yet developers chose the faster tool. This is not Gresham’s law — for most everyday development, “fast enough” simply produces more than “best.”
-
Distribution is the super-lever of the AI era. However good your product is, it loses to being inside the app users already open. ChatGPT’s hundreds of millions of users did as much for Codex as any model improvement.
-
Subsidies are really buying usage data and behavioral habit. The frantic resets were not purchasing user counts — they were purchasing the usage-pattern data of developers building Codex into muscle memory, which feeds model improvement and accelerates the flywheel.
Action Guide by Role
Developers: stop staring at benchmarks. Codex and Claude Code suit different jobs — daily debugging, DevOps, and rapid prototyping favor Codex; complex refactors, code review, and quality-critical work favor Claude Code. Holding both is the correct 2026 engineer posture.
Product people: if you build AI tools, distribution and speed may be scarcer advantages than capability. Kimi K3’s open-source play is another route — open access as distribution to the developer community. Whatever the path, use the RSI logic to judge your own product: is it a pure application, or does it enter the “intelligence producing intelligence” flywheel? If the latter, it deserves cutting everything else for.
Investors: Altman says AI has produced two major product forms — chatbots and coding agents. The third wave, persistent AI agents, is not far off. Start betting now on teams that can enter the “intelligence infrastructure” layer.
Altman’s closing line: “The lag from our early misjudgment on Codex — I can’t afford to make that mistake twice.” If OpenAI is a step late again when the third product form crystallizes, the price next time may be more than Sora and Atlas.
That line is aimed at every AI company.
References: 36Kr, “Codex finally overtakes Claude Code, at heavy cost” · Zhi Dongxi, “Altman’s latest interview” · Artificial Analysis coding-agent evaluation · third-party comparison reports from OpenCode and FireCrawl.
