OpenAI Codex Head Tibo Teases Upcoming Milestone for Codex, Users Speculate Power Reset
Tibo Sottiaux, head of OpenAI Codex, stated that internal dashboards indicate Codex may reach a new milestone tomorrow, reminding users to "keep their Codex safe."
The statement did not disclose specific metrics for the milestone, making it impossible to confirm its relation to active user numbers, task execution volume, code generation volume, paid subscription numbers, revenue, model invocation volume, or product usage duration.
Sottiaux previously revealed that during the unified statistics period for Codex and ChatGPT Work, the two products collectively gained 1 million active users; OpenAI celebrated this milestone by resetting users' weekly usage limits multiple times in advance.
Currently, ChatGPT Plus users share usage limits for Codex and ChatGPT Work: in addition to the weekly limit, a 5-hour usage window restriction has been reinstated; users must wait for limits to refresh or purchase additional credits after reaching either limit.
OpenAI has not yet reinstated the 5-hour limit for the $100 and $200 Pro subscriptions, planning to maintain a more lenient usage arrangement for this tier in the coming months. The tiered limits mean that the growth of Codex users and control of inference costs are directly tied to subscription conversion, limit management, and sales of additional credits.
From a market mechanism perspective, if the new milestone corresponds to an increase in active users or task volume, the primary beneficiaries will be OpenAI's subscription revenue, usage-based billing revenue, and developer ecosystem; however, high-frequency code generation and agent execution will also simultaneously drive up GPU inference, context processing, and tool invocation costs. Plus users are constrained by the usage window, while Pro users retain greater usage freedom, allowing OpenAI to shift high computing demand from a unified subsidy to tiered pricing and paid expansion, with power suppliers, cloud service providers, and data center operators bearing the ongoing increase in infrastructure spending.
Source: Public Information
ABAB AI Insight
The product path of Codex has evolved from an early code completion model to a programming agent focused on task execution. OpenAI released the Codex model in 2021 and used it for GitHub Copilot; since then, the model's capabilities have gradually expanded from generating single code snippets to understanding codebases, editing files, running tests, and suggesting changes. The commercialization experience of GitHub Copilot indicates that the core metrics for development tools are not just the number of registrations, but whether developers integrate the tools into their daily editing, testing, reviewing, and deployment processes; this determines subscription retention and unit user inference costs.
By unifying Codex with ChatGPT Work and reinstating the 5-hour usage window for Plus users, OpenAI is essentially managing limited computing power through product architecture. A unified entry point can enhance user conversion and cross-usage, but it also places text generation, long context analysis, code agent, and tool invocation into the same cost pool; when usage rapidly increases, limits, credits, and Pro tiering become direct mechanisms for allocating inference resources to users with different willingness to pay. The previous early reset of weekly limits to celebrate the addition of 1 million active users indicates that growth milestones can be used as tools for activation and retention, while also temporarily expanding computing subsidies.
This phase can be compared to GitHub Copilot's transition from a personal programming assistant to an enterprise development platform, as well as the shift in cloud computing from fixed packages to usage-based billing. The former increases usage frequency through IDE integration, while the latter matches high-consuming users with revenue through the metering of computing, storage, and network resources. Codex is in a phase of expanding from "model demonstration capability" to "workflow positioning": only when the agent can sustainably complete auditable and reusable engineering tasks will it transform from an auxiliary tool into a stable item in enterprise software budgets; conversely, growth in active users may primarily translate into high-cost free or low-priced inference loads.
Essentially, this is a shift in pricing power: traditional software sells functionality by seat, while AI agents charge based on model capability, context length, task complexity, and tool invocation computing power. As usage depth increases, the true scarce resources are no longer just software licenses, but low-latency inference capacity, reliable agent execution, and enterprise-level workflow integration. Product providers establish new price barriers through limits and credits, allowing infrastructure providers that can reliably obtain computing and power to gain upstream bargaining power; application-layer tools that cannot demonstrate task completion rates and return on investment will face stronger substitution pressure after price transparency.
ABAB News · Cognitive Laws
User growth is revenue, while computing power growth is cost.
Free limits exchange for scale, tiered pricing exchanges for cash flow.
AI does not sell functionality, but replaceable work.