In a major leap for AI-assisted programming, OpenAI has unveiled GPT-5.3-Codex-Spark, a next-generation model designed specifically for real-time coding. The system, now available as a research preview, delivers blazing-fast performance, up to 15 times faster than traditional AI coding models.
Codex Spark is a compact, optimised version of GPT-5.3-Codex, capable of generating more than 1,000 tokens per second on low-latency hardware. This means developers can see changes to their code almost instantly, transforming what was once a back-and-forth with an AI assistant into a seamless, live collaboration.
According to OpenAI, the aim is to make AI programming tools feel “interactive, responsive, and conversational” rather than slow or detached.
OpenAI launched Codex Spark: How does it work
Codex Spark marks the first major outcome of OpenAI’s partnership with Cerebras, announced in January. The model runs on Cerebras’ Wafer Scale Engine 3, a specialised AI accelerator built for extreme-speed inference workloads. This hardware foundation enables Codex Spark to process and respond to developer prompts at near-human speeds.
In practice, the model supports real-time collaboration: developers can interrupt, redirect, or refine the AI’s output as it writes code. By default, Codex Spark performs small, focused edits instead of overhauling entire codebases, keeping the process efficient and under user control. It does not automatically run tests unless explicitly instructed.
Despite its smaller size, the model reportedly performs strongly on software engineering benchmarks, completing coding tasks significantly faster than larger versions. The experience is meant to mimic pair programming, where the AI acts as an instant, tireless co-developer.
Codex Spark currently supports text-only interactions and offers a 128k context window, allowing it to handle extensive codebases with ease. OpenAI describes it as the first in a new family of ultra-fast Codex models that will evolve toward more complex, multimodal capabilities.
How can you access it?
Codex Spark is being rolled out to ChatGPT Pro users via the Codex app, the command-line interface (CLI), and the VS Code extension. During the research preview, the model will have its own rate limits, meaning usage won’t count against standard ChatGPT limits. However, OpenAI has warned that during high demand, users may experience temporary queuing or limited access as the company works to balance reliability.
Looking ahead, OpenAI says Codex Spark is just the beginning of a broader push towards AI tools that combine real-time collaboration with deeper autonomous problem-solving. Future iterations are expected to support larger models, extended context windows, and multimodal input, blurring the line between human and machine creativity in software development.
For developers, Codex Spark may mark the start of a new era, where AI doesn’t just assist coding but codes alongside them, instantly.
