OpenAI introduces GPT-5.3-Codex-Spark for real-time coding

OpenAI has introduced GPT-5.3-Codex-Spark, a coding model designed to respond almost instantly as you write and modify code. The company says it can exceed 1,000 tokens per second, a speed intended for interactive work, not for leaving a task running for hours.

For now, it is a research preview available to ChatGPT Pro users on the latest versions of Codex, OpenAI’s line of coding tools. It is also coming to the API for a small group of partner companies.

A model that codes with you

Larger Codex models are designed for long-running tasks: they can analyze a project, write a lot of code and work for hours with little intervention. Codex-Spark covers the other end of the spectrum: you make a change, see the result and correct it immediately.

For example, you can ask it to adjust a screen’s design, change your mind while it is working or tell it to modify only one specific function. Its default behavior is lightweight: it makes small, targeted changes and does not run tests automatically unless you ask it to.

The idea is to make coding with AI feel less like sending an instruction and waiting, and more like collaborating live.

Speed also depends on infrastructure

Codex-Spark runs on Wafer Scale Engine 3, a specialized processor from Cerebras. OpenAI has integrated this low-latency path into its existing infrastructure so the model can work with Codex without creating a separate system.

The company has also reduced delays between your device, the servers and the model. The announced improvements include:

• 80% less overhead on each round trip between client and server.
• 30% less overhead per token generated.
• 50% less time before the first token appears.

To achieve this, OpenAI added a persistent WebSocket connection, modified parts of its inference system and sped up the way sessions are started. These improvements will eventually reach other models as well.

What it includes and who can try it

Codex-Spark has a context window of 128,000 tokens, meaning it can keep a large amount of text and code available during a session. At launch, it accepts text only, without images or other input formats.

Access is limited to ChatGPT Pro users and has a separate usage limit. During the preview, usage does not count against the standard Codex limit, although queues or restricted access may occur when demand is high.

Availability is gradually expanding to:

• The Codex app.
• The command line, known as the CLI.
• The Codex extension for VS Code.
• The API, initially for some design partners.

OpenAI says Codex-Spark delivers solid performance on SWE-Bench Pro and Terminal-Bench 2.0, two benchmarks for evaluating programming tasks performed by agents. According to the company, it completes those tasks in a fraction of the time taken by GPT-5.3-Codex, although the announcement does not publish detailed scores.

What changes for you

If you use AI to code, the main benefit is not that the model can do entirely new things. It is that you can iterate faster. A style change, a logic fix or a quick test no longer means waiting several seconds each time.

That can be especially useful for:

• Adjusting interfaces while watching the result.
• Exploring several solutions without losing momentum.
• Fixing small errors in large projects.
• Giving new instructions without waiting for a long task to finish.

The tradeoff is that Codex-Spark is optimized for immediate responses, not necessarily for deeper reasoning or longer tasks. OpenAI wants to combine both approaches: a fast mode that works with you and another capable of running long jobs in the background.

The company also plans to add larger models, broader contexts and multimodal inputs, such as images. For now, the important point is the direction: in AI-assisted coding, waiting less can change the experience as much as improving the model’s capabilities. Codex-Spark is OpenAI’s first test of building that interaction in real time.