Tibo Sottiaux: Codex to Accelerate Tenfold This Year
Tibo Sottiaux mentioned that the Codex based on GPT-5.4 already has high speed, but there is still room for acceleration of "at least one order of magnitude" on the engineering side, and this performance leap is expected to be achieved within the year.
This assessment aligns with recent information from various English sources: OpenAI has consistently emphasized engineering optimizations for inference and generation latency in developer updates, while GitHub and several AI programming tools are significantly reducing response times through caching, parallel execution, and model distillation. The industry is shifting from "model capability competition" to "system engineering efficiency competition."
Feedback from some developer communities indicates that the new generation of code models is improving response speed and interaction fluency, which is beginning to change the development process from "waiting for generation" to "real-time collaborative programming."
Source: Public Information
ABAB AI Insight
The core of this information is not about "faster" but about "speed becoming a new variable in production functions." Early competition among large models focused on parameter scale and capability boundaries, but as capabilities become more usable, latency begins to directly determine usage frequency and scene penetration rates. A tenfold speed increase essentially pushes AI from being a "tool" to becoming "infrastructure."
The compounding effects of engineering optimizations are becoming evident. Improvements to the model itself are just the first layer; more crucial are the synergies of inference frameworks, hardware scheduling, caching mechanisms, and task decomposition strategies, all of which together shape the real user experience. Historically, similar paths have appeared in cloud computing and database fields: performance enhancements are often not breakthroughs at single points but rather the result of systematic engineering accumulation.
From an industry structure perspective, this means the focus of competition is shifting from "whose model is stronger" to "whose system is faster and cheaper." Companies with infrastructure capabilities (cloud vendors, model providers) will gain stronger pricing power, while application layer companies that solely rely on model calls will see their marginal advantages compressed.
For the developer ecosystem, the increase in speed changes not just efficiency but also work paradigms. As latency approaches real-time, human-computer interaction shifts from "request-wait-return" to a "continuous collaborative flow," and code generation tools begin to approach the foundational capabilities of IDEs. This is also why judgments like "now is the best time to learn Codex" are emerging—the barriers are lowering, and reliance is deepening.