Credit: Teera Konakan / Getty Images

On Thursday, OpenAI released its first production AI model to run on non-Nvidia hardware, deploying the new GPT-5.3-Codex-Spark coding model on chips from Cerebras. The model delivers code at more than 1,000 tokens (chunks of data) per second, which is reported to be roughly 15 times faster than its predecessor. To compare, Anthropic’s Claude Opus 4.6 in its new premium-priced fast mode reaches about 2.5 times its standard speed of 68.2 tokens per second, although it is a larger and more capable model than Spark.

“Cerebras has been a great engineering partner, and we’re excited about adding fast inference as a new platform capability,” Sachin Katti, head of compute at OpenAI, said in a statement.

Codex-Spark is a research preview available to ChatGPT Pro subscribers ($200/month) through the Codex app, command-line interface, and VS Code extension. OpenAI is rolling out API access to select design partners. The model ships with a 128,000-token context window and handles text only at launch.

The release builds on the full GPT-5.3-Codex model that OpenAI launched earlier this month. Where the full model handles heavyweight agentic coding tasks, OpenAI tuned Spark for speed over depth of knowledge.  OpenAI built it as a text-only model and tuned it specifically for coding, not for the general-purpose tasks that the larger version of GPT-5.3 handles.

On SWE-Bench Pro and Terminal-Bench 2.0, two benchmarks for evaluating software engineering ability, Spark reportedly outperforms the older GPT-5.1-Codex-mini while completing tasks in a fraction of the time, according to OpenAI. The company did not share independent validation of those numbers.

Anecdotally, Codex’s speed has been a sore spot; when Ars tested four AI coding agents building Minesweeper clones in December, Codex took roughly twice as long as Anthropic’s Claude Code to produce a working game.

The coding agent arms race

For context, GPT-5.3-Codex-Spark’s 1,000 tokens per second represents a fairly dramatic leap over anything OpenAI has previously served through its own infrastructure. According to independent benchmarks from Artificial Analysis, OpenAI’s fastest models on Nvidia hardware top out well below that mark: GPT-4o delivers roughly 147 tokens per second, o3-mini hits about 167, and GPT-4o mini clocks around 52.

But 1,000 tokens per second is actually modest by Cerebras standards. The company has measured 2,100 tokens per second on Llama 3.1 70B and reported 3,000 tokens per second on OpenAI’s own open-weight gpt-oss-120B model, suggesting that Codex-Spark’s comparatively lower speed reflects the overhead of a larger or more complex model.

AI coding agents have had a breakout year, with tools like OpenAI’s Codex and Anthropic’s Claude Code reaching a new level of usefulness for rapidly building prototypes, interfaces, and boilerplate code. OpenAI, Google, and Anthropic have all been racing to ship more capable coding agents, and latency has become what separates the winners; a model that codes faster lets a developer iterate faster.

With fierce competition from Anthropic, OpenAI has been iterating on its Codex line at a rapid rate, releasing GPT-5.2 in December after CEO Sam Altman issued an internal “code red” memo about competitive pressure from Google, then shipping GPT-5.3-Codex just days ago.

Diversifying away from Nvidia

Spark’s deeper hardware story may be more consequential than its benchmark scores. The model runs on Cerebras’ Wafer Scale Engine 3, a chip the size of a dinner plate that Cerebras has built its business around since at least 2022. OpenAI and Cerebras announced their partnership in January, and Codex-Spark is the first product to come out of it.

OpenAI has spent the past year systematically reducing its dependence on Nvidia. The company signed a massive multi-year deal with AMD in October 2025, struck a $38 billion cloud computing agreement with Amazon in November, and has been designing its own custom AI chip for eventual fabrication by TSMC.

Meanwhile, a planned $100 billion infrastructure deal with Nvidia has fizzled so far, though Nvidia has since committed to a $20 billion investment. Reuters reported that OpenAI grew unsatisfied with the speed of some Nvidia chips for inference tasks, which is exactly the kind of workload that OpenAI designed Codex-Spark for.

Regardless of which chip is under the hood, speed matters, though it may come at the cost of accuracy. For developers who spend their days inside a code editor waiting for AI suggestions, 1,000 tokens per second may feel less like carefully piloting a jigsaw and more like running a rip saw. Just watch what you’re cutting.