OpenAI Launches GPT-5.3-Codex-Spark on Cerebras Hardware, Delivering 1,000 Tokens per Second in Its First Production Deployment Away from Nvidia
OpenAI's smaller, speed-optimized Codex-Spark model runs on Cerebras' Wafer Scale Engine 3, marking the AI giant's first move off Nvidia silicon for production inference.
Overview
OpenAI released a research preview of GPT-5.3-Codex-Spark on February 12, 2026, a smaller and faster derivative of GPT-5.3-Codex built for real-time coding workflows. The model generates over 1,000 tokens per second — roughly 15 times faster than the standard GPT-5.3-Codex — by running on Cerebras’ Wafer Scale Engine 3, according to OpenAI’s announcement. The launch marks the first production milestone of a multi-year partnership between OpenAI and the chip maker, and the first time OpenAI has deployed a GPT model on non-Nvidia hardware for production inference.
What We Know
Codex-Spark is a distilled, text-only model with a 128K context window, designed to prioritize interactive speed over the deep reasoning of its larger sibling. It is available as a research preview to ChatGPT Pro subscribers through the Codex app, command-line interface, and VS Code extension, with API access extended to select design partners, as reported by Help Net Security.
OpenAI describes the model as more capable than GPT-5.1-Codex-mini while completing tasks in a fraction of the time compared to the full GPT-5.3-Codex, according to Cerebras’ blog post. On SWE-Bench Pro, a benchmark evaluating agentic software engineering capability, Codex-Spark reaches near the same accuracy as the full-size model at dramatically higher speed. The trade-off surfaces on Terminal-Bench 2.0, which tests complex multi-step terminal operations: the full GPT-5.3-Codex scores 77.3 percent while Codex-Spark achieves 58.4 percent, as noted by Techzine. Both figures exceed the prior generation GPT-5.2-Codex’s 64 percent.
OpenAI framed the model as part of a dual-mode vision for Codex. According to the company, Codex-Spark represents “the first step toward a Codex that works in two complementary modes: real-time collaboration when you want rapid iteration, and long-running tasks when you need deeper reasoning,” as reported by Dataconomy.
Infrastructure Overhaul
Beyond the model itself, OpenAI announced significant changes to its serving infrastructure. The company introduced a persistent WebSocket connection that keeps a live channel open between client and server, eliminating the overhead of initiating a new request for each interaction. Combined with optimizations to the Responses API, the changes reduced per-roundtrip overhead by 80 percent, per-token overhead by 30 percent, and time-to-first-token by 50 percent, according to OpenAI. The WebSocket path is enabled for Codex-Spark by default and is expected to become the default for all OpenAI models.
The Cerebras Partnership
The hardware underpinning Codex-Spark is the Cerebras Wafer Scale Engine 3, a purpose-built accelerator containing 4 trillion transistors and the largest on-chip memory of any AI processor, according to Cerebras. Cerebras’ architecture scales out to thousands of systems, extending memory capacity into the multi-terabyte range to support trillion-parameter models.
The deployment stems from OpenAI’s multi-year agreement with Cerebras, announced in January 2026, securing up to 750 megawatts of computing power over three years, as reported by Techzine. TechCrunch characterized the launch as “the first milestone” in that relationship. The arrangement is complementary rather than a wholesale replacement: Nvidia GPUs remain OpenAI’s foundation for training and broad inference, while Cerebras handles workloads requiring ultra-low latency.
What We Don’t Know
Pricing for Codex-Spark remains unannounced. Simon Willison noted that it is “not yet clear what the pricing will look like,” a significant open question for developers evaluating the model for production use. It is also unclear when or whether the research preview will graduate to general availability beyond ChatGPT Pro.
The model carries independent rate limits during the preview period, and usage does not count toward standard ChatGPT rate limits, according to Help Net Security. Whether those favorable terms survive a broader rollout is an open question.
OpenAI has not disclosed the exact size of the distilled model relative to the full GPT-5.3-Codex, nor has it published detailed safety evaluations specific to Codex-Spark beyond confirming it underwent baseline safety processes.
Analysis
Codex-Spark represents a meaningful strategic shift on two fronts. First, it validates specialized silicon for AI inference at production scale. Cerebras has demonstrated high token throughput before — it ran Llama 3.1 70B at 2,000 tokens per second in October 2024 — but deploying an OpenAI flagship model on non-Nvidia hardware is a first for the industry’s most prominent AI company. The signal to the broader AI chip market is significant: purpose-built inference accelerators can earn a seat alongside Nvidia GPUs in frontier AI infrastructure.
Second, the dual-mode vision for Codex — fast, interactive Spark for real-time iteration alongside the slower, deeper full model for autonomous tasks — reflects a maturing understanding of how developers actually work with AI coding tools. Speed matters for maintaining flow state during iterative development; raw capability matters for complex, multi-step engineering tasks. Shipping both as complementary options within the same product line acknowledges that no single model optimally serves both needs.