OpenAI Codex Head Tibo: Pre-release Code Freeze Has Basically Disappeared
Thibault Sottiaux, who is responsible for OpenAI's Codex and ChatGPT core products, posted that the pre-release code freeze has basically disappeared; in the future, code may even be generated online upon request and under constraints.
Sottiaux, known publicly as Tibo, currently oversees core products and platforms, covering both Codex and ChatGPT, and reports to Greg Brockman. He has become a public face in the developer community for frequently resetting Codex usage limits after failures. A recent interview with Pragmatic Engineer clarified that he was involved in creating Codex, with the CLI written in open-source Rust, aiming to run proxy instances on massive cloud hosts.
Codex is reported to be one of OpenAI's fastest-growing revenue lines. The product path is an open-source orchestration layer that can connect to multiple models, rather than being locked to proprietary weights; the team has publicly supported linking subscription limits to other programming proxy tools. It continues to compete with Anthropic's Claude Code in the separation of shell and model: the same model can be swapped with a different shell, allowing benchmarks to be rewritten.
"Code freeze" refers to the discipline of freezing submissions during traditional release windows, only fixing blocking defects. Tibo describes it as outdated, effectively acknowledging that proxies have compressed patches, regressions, and release notes into the same pipeline. The next step of "on-demand online generation" transforms repositories from pre-written texts into instant products under constraints—constraints becoming a harder interface than branching strategies.
This is not a new feature announcement, but a philosophy of release: static snapshots give way to regenerable code. Safety and reproducibility will shift from "a certain git hash" to "can a certain set of constraints regenerate the same behavior?"
In market mechanisms, buyers are enterprises and individuals writing millions of lines of patches daily using proxies, while sellers are model companies that incorporate generation, testing, and release into the same quota; funds flow from ChatGPT's paid tiers to tokens and cloud execution. Beneficiaries are orchestration layers that can turn constraints into products, while those under pressure are traditional release managers who rely on frozen windows to control quality and compliance audits that depend on fixed binaries. The event-driven nature comes from the penetration rate of proxy coding, rather than a specific version number.
No exchange data available. What can be verified is the spokesperson's position and the existence of open-source repositories, not a census of "freezes have disappeared in all companies."
Source: Public Information
ABAB AI Insight
Tibo has dismantled the oldest brakes in software engineering. The premise of freezing is that humans write code, humans review code, and humans fear going live on Fridays; proxies compress writing and testing to minutes, making the window itself a delay. On-demand generation transforms the "repository" from an asset into a function: input constraints, output runnable text. OpenAI, with its open-source Rust shell, is competing for the scheduling rights of this function, rather than the copyright of a specific file.
The capital path is for subscription limits to become a continuous compilation budget. Whoever controls the constraint language controls the certification of regenerable code. The interlocking of Claude Code and Codex models indicates that models are interchangeable, and the shell and strategy layer are where the rent lies. Resetting limits after failures turns computational waste into customer relationships, also indicating that the cost unit of generative releases is now tokens rather than man-hours.
Similar structures appear in JIT compilation replacing pre-compilation, infrastructure as code replacing data center manuals, and content websites rendering pages on demand. The industry is in transition: tools can already generate, while organizations still use branch protection as proof of quality. Audits, licenses, and vulnerability disclosures still assume a frozen tree exists.
Structural judgment belongs to technological substitution. The mechanism is generative proxies replacing human gatekeeping within frozen windows. Once the substitution is complete, releases shift from "code at a certain moment" to "real-time solutions under certain constraints"; quality transitions from prohibiting submissions to whether constraints can be reproduced, replayed, and held accountable.
ABAB News · Cognitive Law
Where code freeze disappears, constraints will become the new version number
On-demand generation transforms the repository from an asset into a function
Whoever defines the constraints defines what counts as a successful release