The real question is, of the 6 million, how many are one-time users versus how many are truly dependent.

Author and source: 0x9999in1, ME News



TL;DR

• The combined active users of Codex and ChatGPT Work have reached 6 million. Codex’s weekly active users grew from under 1 million in February to 5 million by early June, and then to 6 million by mid-July—achieving in five months what most SaaS companies take three years to accomplish.
• OpenAI has temporarily removed the 5-hour usage limit for Plus, Business, and Pro plans and reset all quotas uniformly. This isn't generosity—it's a product manager doing the math.
• The usage calculation for GPT-5.6 Sol has been reduced, consuming less quota for similar tasks. The exact reduction percentage has not been disclosed, but it provides a tangible extension of usage for high-frequency users.
• Codex’s growth curve was steeper than ChatGPT’s in its early days; but the competition it faces is far tougher—Cursor, Claude Code, and Windsurf are all vying for the same water in the same river.
• OpenAI's intent is clear: to stake a claim in developers' minds and make Codex the default choice in the IDE era.
• But growth rate doesn't equal victory. The real question is: of the 6 million, how many are "one-time users," and how many are "hooked"?

I. What does the number 6 million actually mean?

To cut to the chase: Codex's growth rate has entered OpenAI's internal category of "must-重点押注" initiatives.

When Tibo mentioned the number 6 million, his tone was calm. But when viewed over the timeline, this number is anything but ordinary.

In February, OpenAI released the Codex desktop version, with fewer than 1 million weekly active users. By early June, Axios reported that weekly active users had surpassed 5 million, more than fivefold. By mid-July, the official figure increased to 6 million combined active users for Codex and ChatGPT Work.

Over a month, an additional one million.

What does this mean? Anyone who has worked in SaaS knows that growing a product from 0 to 1 million relies on novelty and the early-mover advantage. Growing from 1 million to 5 million depends on product strength and reputation. Beyond 5 million, every new user must be acquired with real money.

The current curve of Codex doesn't appear to have entered the phase where "each step is expensive."

Why? Because it sits atop two massive sources of potential energy: the brand spillover from ChatGPT itself and the ongoing capability spillover from the GPT-5 series of models. Together, these two forces enable Codex to spread organically within the developer community with little need for independent user acquisition.

But there’s an important question to clarify here: 6 million is "active," not "paying" or "heavily using."

OpenAI has not disclosed the DAU/WAU ratio or retention curves. Of the 6 million, how many open the platform only once per week? How many have truly integrated Codex into their daily workflows? We don’t know.

So this number looks good, but it's not enough to draw a conclusion.

II. Canceling the 5-Hour Limit: Generosity or Calculated Move?

The most easily misunderstood aspect of Codex's recent action is the "temporary removal of the 5-hour usage limit for Plus, Business, and Pro."

Many people’s first reaction: OpenAI is being generous.

But if you've ever worked on an internet product for even a day, you know—limits are never about generosity or stinginess; they're about the product manager doing the math.

The previous 5-hour limit was set due to GPU scarcity, high per-task costs, and the need to protect the experience of heavy users from being exploited. We can now remove it for several reasons:

First, the reasoning cost for GPT-5.6 Sol has decreased. This has prompted a corresponding adjustment in usage calculations—similar tasks now consume less quota. The exact reduction percentage hasn't been disclosed, but the direction is clear: the computational cost per task is being reduced.

Second, Codex’s competitors are applying pressure. Cursor has over a million paid users, and Anthropic’s Claude Code has earned a strong reputation among developers. If OpenAI continues to restrict high-frequency users with a five-hour limit, it risks pushing its most loyal users toward competitors.

Third, and most crucially—OpenAI isn’t seeking to maximize every penny of profit; it’s focused on securing developers’ minds first.

This is a classic platform strategy: first, make you love it—so much so that you can’t do without it, where all your project code is generated by Codex and your entire team’s workflow revolves around it. At that point, raising prices, imposing restrictions, and launching an enterprise version all become natural next steps.

Amazon’s original Prime free shipping followed this logic; Netflix’s early unlimited streaming followed this logic; now it’s Codex’s turn.

So don’t be moved by “removing restrictions.” It’s not a gift—it’s an investment.

Three Real Drivers of Codex Growth

Setting aside the narrative, what exactly enabled Codex to achieve 6x growth in five months?

The GPT-5 series represents a generational leap in code-related tasks compared to the GPT-4 era. This is not marketing rhetoric—it’s the result of developers voting with their actions.

A clear indicator: On real-world software engineering benchmarks like SWE-bench, the pass rate for the GPT-5 series has significantly improved over GPT-4o. Combined with optimizations in the reasoning chain by GPT-5.6 Sol, tasks that AI previously struggled with—such as handling long contexts, cross-file refactoring, and complex debugging—are now performed convincingly.

When a model improves by one level, its user base expands by one circle. This logic is more effective for AI products than any growth hacking tactic.

Codex is no longer just a "web product"—it's a comprehensive suite of tools embedded into developers' workflows.

The desktop version caters to developers accustomed to IDE interactions, the CLI targets seasoned hackers who live in the terminal, and the web version handles occasional use cases.

Standing on three legs is much more stable than standing on one.

This point is the most easily underestimated.

For ordinary developers, the cost of deciding whether to try Codex is far lower than deciding whether to try Cursor. Why? Because they are already using ChatGPT—the account is ready, the payment is set up, and the trust is already there.

OpenAI has integrated Codex into its Plus, Pro, and Business subscriptions, effectively making every paying ChatGPT user a potential Codex user.

This is a distribution advantage that no amount of funding from Cursor can buy.

Four, but Codex is not without concerns

At this point, focusing only on the growth story would be too thin.

Codex's true competitor isn't Cursor in numerical terms, but the erosion of developers' trust in the very concept of AI programming tools.

Anyone who has used AI to write code understands the experience: the first ten minutes are great, but the next hour makes you question reality. The generated code looks right, but it doesn’t run correctly. Fixing one bug introduces three new ones. Ask it to refactor, and the entire file’s style changes completely.

Does Codex have this experience? Yes. Does Cursor have it? Yes. Does Claude Code have it? Yes, as well.

So, the key to winning this battle is determining what percentage of the 6 million active users truly become "daily users."

OpenAI's adjustment to the usage calculation for GPT-5.6 Sol is on the right track, but since they haven't disclosed the extent of the reduction, it suggests they may still be testing internally what constitutes sufficient usage.

Here’s another uncomfortable truth: By the second half of 2025, Cursor’s funding valuation had surged to nearly $10 billion, while Anthropic’s Claude Code, leveraging the stability of its Claude 4 series models on long tasks, is steadily capturing market share in the enterprise sector.

Codex's 6 million was achieved in such a crowded market. Running fast doesn't mean running far.

V. What the Algorithm Adjustments of GPT-5.6 Sol Reveal

Alongside this user growth release, we are also adjusting the GPT-5.6 Sol usage calculation.

This appears to be an operational move, but it's actually a technical signal.

OpenAI's willingness to "reduce the quota consumed by similar tasks" implies two things are simultaneously true: first, inference efficiency on the model side has improved, and second, resource allocation on the scheduling side has been optimized.

The former is due to the model architecture, and the latter is due to the infrastructure. Only when both are combined can we confidently tell users, "You can use more."

This also indirectly answers a question the industry has been asking: Can the inference cost of the GPT-5 series be reduced to the level of GPT-4?

From this adjustment, the answer is getting closer to "yes."

Every 30% reduction in inference cost creates additional pricing flexibility and shifts the competitive landscape—that’s what makes this adjustment truly significant.

OpenAI has not disclosed the exact magnitude of the reduction. This vagueness itself is a strategy—it leaves room for user imagination and provides flexibility for future adjustments.

Six: How will the second half of this war be fought?

Codex reaching six million is just the end of the first half. The keywords for the second half, in my view, are three: corporatization, multimodality, and agentization.

ChatGPT Work has been merged with Codex for active user calculations—the signal is clear: OpenAI is pushing Codex toward the enterprise market. Enterprise customers have an ARPU more than ten times higher than consumers and much higher retention rates. This is the path all SaaS companies must take.

But the enterprise market is a tough nut to crack. Compliance, on-premises deployment, access management, audit logs, and SSO integration—each of these is something consumer products don’t need to consider. Whether OpenAI is ready for this remains to be seen in the coming quarters.

Code is more than just text. Charts, screenshots, design mockups, whiteboard photos, and terminal outputs can all serve as inputs for programming. If Codex could truly integrate GPT-5’s multimodal capabilities into the coding workflow—such as “Write the interface based on this Figma design” or “Tell me how to fix this based on this error screenshot”—that would be the real differentiator.

There haven’t been any major moves from OpenAI in this area yet. But since the model capabilities are already there, productization is just a matter of time.

Codex currently primarily operates on a "you speak, I write" interaction model. The next step will certainly be "you speak, I do"—autonomously completing entire tasks, making changes across multiple files, automatically running tests, and automatically submitting pull requests.

OpenAI previously released the Codex Agent for testing in this direction, but the current experience still falls short of being truly ready to let it operate independently.

If we succeed in agentization, Codex’s moat will deepen. If we fail, it will forever remain just an “advanced version of GitHub Copilot.”

Seven: Three Calm Judgments for Users

After all this, it ultimately comes down to one question: How should ordinary developers view Codex’s recent growth and adjustments?

First, cancel the 5-hour limit—it’s worth taking advantage of. This is OpenAI’s genuine opportunity to win your loyalty with real money; don’t hesitate.

Second, don’t treat the 6 million figure as a definitive winner. The AI programming space is still in its early stages—the landscape could shift multiple times within a year. Someone using Codex today might switch to Cursor tomorrow, and then try Claude Code the day after. Tools are just tools—don’t bet everything on any one of them.

Third, what truly determines your efficiency isn't which tool you use, but whether you know how to use it. The same Codex can boost someone’s productivity threefold, while another person spends their days debugging until they’re exhausted. This gap has little to do with the tool itself.

Eighth, in conclusion

OpenAI dismisses the figure of 6 million as if it were trivial. Yet every user on this curve is casting their time and trust as a vote for an undefined future.

Will Codex win? I don't know.

Will Cursor lose? I don't know.

But one thing is certain: the battle over AI programming has only just begun. Whichever model is stronger, whose product is more intuitive, and whose pricing is more aggressive—every variable is still in flux.

The buzz is theirs; the code is yours.

At this point, I feel these words should be sent to every developer using Codex, Cursor, or Claude Code—no matter how powerful the tools are, they can’t replace your judgment at the moment you hit enter.

Six million is just the beginning. The real story is yet to come.

Reference materials

1. Axios, "OpenAI's Codex weekly active users surpass 5 million," reported in early June 2026
2. Official product update announcement from OpenAI: Tibo's statement regarding the combined active users of Codex and ChatGPT Work, 6 million, July 2026
3. Announcement of OpenAI Codex Desktop Version, February 2026
4. OpenAI product notice regarding adjustments to GPT-5.6 Sol usage calculations and temporary removal of the 5-hour usage limit for Plus, Business, and Pro, July 2026
5. Cursor (Anysphere) 2025 H2 funding and valuation-related media coverage
6. Anthropic Claude Code product page and capabilities of the Claude 4 series models
7. SWE-bench benchmark public leaderboard, featuring performance data of the GPT-5 series on real-world software engineering tasks.