If you've been using Codex in ChatGPT Pro recently, you might have noticed a new option called Codex-Spark with response speeds that feel almost unreal.
This isn't the result of parameter tuning — the underlying hardware has changed.
What chip is it running on
GPT-5.3-Codex-Spark runs on the Cerebras WSE-3 (Wafer Scale Engine 3), not Nvidia GPUs.
This is the first time OpenAI has deployed a formal production-grade model on inference hardware outside of Nvidia.
The WSE-3 is a silicon wafer roughly the size of a dinner plate, integrating over 4 trillion transistors with massive on-chip memory, designed to eliminate latency bottlenecks caused by data movement.
The real-world result: over 1,000 tokens per second.
A standard GPT-5.3-Codex runs at about 65 tokens per second. Codex-Spark is roughly 15 times faster.
The scale of the OpenAI-Cerebras deal
OpenAI signed a multi-year contract with Cerebras in January this year. The commitment: phased deployment of 750 megawatts of Cerebras compute power by 2028, with a total contract value exceeding $10 billion.
But Sam Altman has not dismissed Nvidia either, stating that "Nvidia makes the best chips in the world" and that the long-term partnership remains unchanged. OpenAI's current strategy is multi-vendor: Nvidia for training, Cerebras for low-latency inference, AMD with a 6 GW agreement, and Broadcom custom chips in development.
From Cerebras' side, Sachin Katti said: "Bringing wafer-scale computing into production gives us a new way to keep Codex responsive in latency-sensitive tasks."
What Codex-Spark can be used for
OpenAI positions it as a "daily productivity tool for developers" — not for deep complex reasoning, but for work that requires rapid iteration and instant feedback:
- Quick code editing and local refactoring
- Real-time debugging and error localization
- Writing PRDs, user research documents, and monitoring metrics
- Managing tests and tracking task progress
The core value is no waiting. Modify a function, ask a question, test a piece of logic — near-instant response.
It supports an "interruptible" workflow — you can interrupt the AI's execution at any time and adjust direction, rather than waiting for it to finish a long block before seeing results.
Benchmark data
| Metric | GPT-5.2-Codex | GPT-5.3-Codex-Spark |
|---|---|---|
| Terminal-Bench 2.0 | 64% | 77.3% |
| Inference speed | ~65 token/s | 1000+ token/s |
| Relative output speed | Baseline | ~25% faster |
| Token consumption (some tasks) | Baseline | ~50% less |
Benchmark scores improved while speed increased 15x — two things that rarely happen together.
The bigger picture
For now, Codex-Spark is only available to ChatGPT paid users. API access is still on the way, with no separate pricing announced.
OpenAI revealed that Codex now has over 1 million weekly active users. At this scale, inference latency directly determines product experience — it's not just a number on a spec sheet.
The more significant signal: when Cerebras goes from "theoretically able to replace Nvidia" to "the first non-Nvidia hardware OpenAI puts into production," that shift carries real weight in the chip industry.
As for whether Nvidia will be affected — on the training side, no one can challenge them in the short term. But in the inference market, this crack has already opened. And it was opened by OpenAI.
Sources: OpenAI launches GPT-5.3-Codex-Spark on Cerebras chips — marks AI giant's first production deployment away from Nvidia (Tom's Hardware); Introducing OpenAI GPT-5.3-Codex-Spark Powered (Cerebras AI Blog); CocoLoop; OpenAI's rapid GPT-5.3-Codex model moves beyond simple coding tasks (SiliconANGLE)