OpenAI Codex Project Lead: Future Models Will Be Better at Code Deletion and Simplification
Thibault Sottiaux, the lead of the OpenAI Codex project, stated that future models will be significantly better at deleting code and simplifying it.
The statement did not provide model names, launch dates, or code deletion ratios. He focused on two capabilities: writing less and shortening existing code. This statement addresses the amount of code being generated by programming agents, rather than a list of new features.
Sottiaux publicly identifies as the head of the Codex project, which includes about 30 research and product engineers, covering command line, editor plugins, and cloud agents. Almost all software engineers at OpenAI use Codex, and code reviews primarily go through this agent. In previous interviews, he referred to complex scaffolding as something to cope with rather than expand, advocating for gradually removing external processes as models become stronger.
This product line has already undergone a reduction. In May 2026, he announced that to simplify Codex cluster management, gpt-5.2 and gpt-5.3-Codex would be removed from Codex logged in via ChatGPT starting June 2. The default model for the free plan was changed to gpt-5.5, while the old models remain in the API. The previous reduction involved optional models, not the lines of code in user repositories.
He also mentioned in another context that maintenance costs have significantly decreased compared to two or three years ago, with review, deployment, and regression interception being as automated as possible, allowing individual changes to be pushed to about 1 billion users on the same day. If the code deletion capability keeps pace, it will target the volume and branching complexity left after rapid integration.
This is an expectation of capability, not an order or pricing announcement. The buyers are teams that have already integrated Codex into continuous integration and repositories, while the sellers remain OpenAI's models and computing power. The event is driven by code bloat approaching maintainability limits. Beneficiaries are small teams that can delegate refactoring and deletion to agents, while those relying on counting lines for pricing and still building scaffolding tools are under pressure. No new funding or rate disclosures were made.
Source: Public Information
ABAB AI Insight
Sottiaux's path is from DeepMind and Google to OpenAI, subsequently leading about 30 people to make Codex the default programming partner internally. The public narrative has always been about dismantling scaffolding: external processes will remain until the model can stand on its own, then they will be removed. The removal of gpt-5.2 and gpt-5.3-Codex from Codex logged in via ChatGPT in May 2026, leaving only the API, is already a simplification of the cluster using model menus, not a marketing simplification.
The capital is not in acquiring codebases, but in harvesting computing power from "multi-model parallel supply" into fewer cutting-edge defaults. gpt-5.5 has become the default for the free plan, while the old Codex models exit this entry point, saving on service fleet and routing complexity. The motivation is that as agents write faster, repositories grow fatter, and the next step must ensure the same model is responsible for deletion and merging; otherwise, the internal release pace for about 1 billion users will be bogged down by maintenance debt.
This can be compared to the early completion path of GitHub Copilot, which still relies on long prompts and multi-layer toolchains to support agents. Copilot initially sold "write one more line," while Codex has reached a point where "leave one less line" can accommodate full repository refactoring. The industry position is in a control phase after expansion: agents have entered continuous integration, and the difference shifts from whether they can write to whether they can delete without altering behavior.
Structurally, this is the second step of technological substitution. The first step replaced typing, and the second step replaces manual refactoring. The mechanism is that as generation costs approach zero, the scarce resources become readability and regression risk. Whoever first makes deletion a verifiable action will take the maintenance budget. The pricing model that values writing more will be the first to be altered by this step.
ABAB News · Cognitive Laws
Once generation becomes cheap, deletion starts to become valuable.
Scaffolding is a transitional tax, not a moat.
Those who write quickly will get rich first, while those who can delete will remain.