Codex's growth relies on locking in users, not technological superiority—OpenAI is betting on migration costs.

Behind Altman's Numbers: Trading "Reset" for Lock-in

Sam Altman posted that Codex has 3 million weekly active users. What's noteworthy isn't the number itself, but the approach: high-frequency resets limit the burn of computing power to buy time, aiming to deposit developer code context into Codex before Anthropic's Claude Code penetrates deeper.

• In simple terms, this is "defensive expansion," not natural growth driven by a better product.
• Official statements point toward 10 million future users. Each reset pushes the team to bind more context, dependencies, and workflows to Codex.
• Once context becomes an asset, migration costs emerge, and relying solely on model capabilities makes it hard to replicate this stickiness.

Recent research also exposes structural issues:

• Tool weekly active penetration reaches 95%, but only 29% of engineers dare to put AI-generated code directly into production.
• Heavy usage and trust are mismatched, indicating that actual quality and risk control haven't caught up, and surface growth masks underlying risks.

Market opinions on this wave of "reset to acquire users" vary:

• Optimists estimate about 30 million developers on GitHub, believing it has already penetrated 10% globally.
• Cost skeptics question how long a monthly $100–200 compute expense per user can sustain.
• Community feedback shows large fluctuations in reset limits, suggesting infrastructure may be strained.

All signals point to the same conclusion: growth is about "抢位" (seizing position), not "领跑" (leading).

• Claude Code is ahead in preference: 46% of survey respondents say it's "most popular," while Copilot is only 9%; Microsoft's ecosystem reputation and version fluctuations cause negative migration. Codex is still catching up.
• Asynchronous parallel vs. IDE real-time: Codex's multi-task agent excels at parallel capabilities, but in daily development, human-AI on-screen interaction and real-time feedback can't match Cursor's IDE experience; mid-term, they will likely be used together rather than one dominating.
• Trust gap as hidden cost: 84% of developers are using it, but low trust increases rework and validation costs, which capital markets may underestimate.

The "Burning Compute" stance more resembles passive defense

Discussions around resets boil down to two points:

• Is this a defensive move against Claude Code's roughly $2.5 billion annual volume?
• Or is it masking OpenAI's commercial pressure?

External signals fill in the story:

• Pragmatic Engineer's research shows Claude reached peak preference in 8 months;
• NxCode's comparison indicates Copilot's context window is limited, while Codex has an advantage in parallel processing.

The industry’s main narrative has shifted from "AI coding has been conquered" to "moats come from context and integration, not just raw compute." Companies with strong integration and high workflow stickiness have an edge.

Policy concerns about AI energy consumption are background risks, but underestimated is the "lock-in speed": once user scale hits 10 million, even if not highly profitable, a strong path dependence could form.

| Camp | What they see | How it influences judgment | My view | |------|----------------|----------------------------|---------| | Growth advocates | 3 million weekly active users, accelerated by resets | Seen as leading expansion, boosting developer tool valuation | Overestimated. Growth is defensive, not natural retention; lock-in requires trust to rebound first | | Cost skeptics | $100–200 monthly compute per user, resets are money-burning | Scale and marginal limits suppress expansion expectations | Real risks exist. Time is traded for space, but profits are eroding, benefiting more cost-efficient competitors | | Competitor watchers | Claude preference 46%, Copilot declining | Focus on autonomous agents and trust erosion in old tools | Mixed use is better mid-term. Reset underestimates the importance of real-time daily interactions | | Trust advocates | 95% are using it, but only 29% dare to deploy in production | Verification and governance become bottlenecks, moving toward human-AI hybrid | Underestimated. Low trust limits enterprise scaling; rework costs are not fully accounted for |

All signals point to the same conclusion: OpenAI's position isn't stable, but they are betting aggressively—core strategy is "turning context into a moat."

Summary: This reset strategy aims to turn Codex into a "lock-in tool," rather than relying on "model superiority." Anthropic holds an advantage in enterprise trust. For developers, a more realistic path is mixing Codex + Cursor for efficiency; for investors, watch whether compute costs can be sustained.

Importance: High
Category: Industry trend, developer tools, market impact

Conclusion: It’s late to treat this as a "sure-win" now; the beneficiaries are developers quickly adopting hybrid workflows and funds capable of enduring compute costs and planning long-term. Short-term traders have no advantage, nor do long-term holders betting solely on one tool stack.