OpenAI Codex-Spark writes code up to 15 times faster

What is new

OpenAI has launched GPT-5.3-Codex-Spark, a new and lighter version of its coding model GPT-5.3-Codex. It is built for one thing: responding almost instantly while you program. OpenAI says it writes code about 15 times faster than previous versions. That equals more than 1,000 tokens per second. Tokens are the small pieces of text an AI model reads and writes in.

The model does not run on the usual Nvidia chips. Instead, OpenAI uses chips from Cerebras (the WSE-3 model), which are the size of a dinner plate. This is part of a major partnership between the two companies.

Spark is for now a so-called research preview, meaning an early trial version. It is open to ChatGPT Pro users in the Codex app, in the command-line tool and in the VS Code extension. A small selection of partners can also use it via the API, and OpenAI will give more people access over the coming weeks.

What is clever

The advantage is speed. When an AI tool responds in a split second, you can work with it like a colleague sitting next to you: fix a small thing, see the result and fix it again without waiting. It is good for quick experiments and small adjustments.

The model has a context window of 128,000 tokens. That is the amount of text it can hold in its head at once, which equals a decent pile of code. It is text-only, so it cannot understand images. During the preview, it has its own usage limits, so usage does not count toward the regular quota in the subscription.

Cheaper or better?

Spark is faster, but not smarter than its big brother. The large GPT-5.3-Codex is still the model for heavy and complex multi-step tasks where the model needs to think deeply. Spark is a smaller model that trades some of that depth for speed.

On price, the picture is unclear. For ordinary users, access requires a ChatGPT Pro subscription, and I have not been able to confirm an official monthly price for this particular model. Some third-party sites list API prices of $1.75 per million input tokens and $14 per million output tokens, but this has not been confirmed by OpenAI in the material I have been able to read, so the figures should be taken with caution. No independent measurements against competitors such as Anthropic's Claude Code or Google's Gemini have been found either.

What it is good at

Spark is especially good for developers who want answers in real time: quick fixes, small changes to existing code, prototypes and playing with ideas. It is less suited to large, complex projects where quality and thoroughness weigh more than speed.

For ordinary employees outside IT, the news is mostly a signal that AI coding tools are becoming both faster and more interactive. Whether and when Spark will come to Denmark and the EU outside the Pro subscription has not yet been announced.