OpenAI Head Romain Huet: Codex Integrated into Main Model System
Romain Huet, head of developer experience at OpenAI, revealed in a response to users on the X platform that starting from GPT-5.4, the previously independently maintained Codex programming-specific model series has been merged into a single system with the main model and will no longer iterate separately. GPT-5.5 significantly enhances the capabilities of agent-driven autonomous programming and computer use (directly controlling computers to complete tasks).
This merger means developers no longer need to switch between general models and programming models; a single model will support dialogue, complex reasoning, and deep programming tasks simultaneously. Codex, previously an independent series, had long served code generation and completion scenarios, and is now fully integrated into the o series backbone.
This change will directly reshape developer workflows, reduce model selection costs, and promote the implementation of agentic applications, especially creating a stronger closed loop in automated coding and computer operation scenarios.
Source: Public Information
ABAB AI Insight
Romain Huet's team at OpenAI has previously pushed for model unification, similar to the gradual alignment of the early code-davinci series with GPT-4o in 2023-2024. The complete merger in GPT-5.4 is a continuation of this strategy. OpenAI has historically invested heavily in dedicated training data and fine-tuning resources for Codex, and the decision to integrate it into the main model indicates that they believe the scaling of the backbone model is sufficient to cover professional programming needs.
From a capital and business perspective, the merger allows OpenAI to significantly reduce the computational and engineering costs of maintaining multiple models in parallel, concentrating resources on the iteration of a single cutting-edge model. This approach is similar to Google's unification of Bard/Bard Advanced into the Gemini mainline, with the core logic being to accelerate the commercialization of agentic products by reducing fragmentation, aiming to encourage paid developers and enterprise clients to subscribe long-term to a single high-performance entry point.
Historically, Anthropic's Claude series also underwent a transition from independent code models to a unified multimodal system, ultimately gaining market share in the enterprise coding sector. OpenAI's move now positions it ahead of Claude 4 and Gemini 2.5 in the agentic coding field, establishing a competitive advantage in "full-stack coverage with a single model," especially creating short-term barriers in computer use capabilities.
This news essentially pertains to a restructuring of the industry chain: OpenAI is transforming the programming toolchain from a parallel structure of "dedicated models + general models" to a vertically integrated structure of "single super model + tool invocation." This change is driven by the fact that the marginal returns of scaling law still exceed those of dedicated optimization, while the rise of agentic and computer use demands a unified contextual understanding capability from the model.