OpenAI has achieved a significant milestone in model efficiency with the optimization of GPT-5.3-Codex-Spark. According to Thibault Sottiaux, a researcher at the organization, the model now operates 30% faster than previous iterations, pushing the boundaries of real-time code generation and reasoning capabilities.

The primary highlight of this update is the impressive throughput, with the model now serving at over 1200 tokens per second. For context, tokens are the basic units of text—like syllables or word fragments—that AI models process and generate. High token speeds are crucial for applications requiring near-instantaneous feedback, such as interactive coding assistants or complex agentic workflows where latency acts as a primary bottleneck for user experience.

This development suggests that OpenAI is focusing heavily on the Codex lineage, likely refining how the model handles structured data and programming logic. By increasing the speed of inference (the process where a trained model generates a response), developers can expect more fluid interactions and lower operational costs for large-scale deployments. The Spark suffix likely indicates a lightweight version of the primary architecture, optimized specifically for high-velocity output without sacrificing core logic.

OpenAI has achieved a significant milestone in model efficiency with the optimization of GPT-5.3-Codex-Spark. According to Thibault Sottiaux, a researcher at the organization, the model now operates 30% faster than previous iterations, pushing the boundaries of real-time code generation and reasoning capabilities.

The primary highlight of this update is the impressive throughput, with the model now serving at over 1200 tokens per second. For context, tokens are the basic units of text—like syllables or word fragments—that AI models process and generate. High token speeds are crucial for applications requiring near-instantaneous feedback, such as interactive coding assistants or complex agentic workflows where latency acts as a primary bottleneck for user experience.

This development suggests that OpenAI is focusing heavily on the Codex lineage, likely refining how the model handles structured data and programming logic. By increasing the speed of inference (the process where a trained model generates a response), developers can expect more fluid interactions and lower operational costs for large-scale deployments. The Spark suffix likely indicates a lightweight version of the primary architecture, optimized specifically for high-velocity output without sacrificing core logic.