GPT-5.3-Codex-Spark: OpenAI's First Real-Time Coding Model on Cerebras Hardware

OpenAI Breaks New Ground with Real-Time Coding

On February 12, 2026, OpenAI unveiled GPT-5.3-Codex-Spark, a smaller, speed-optimized variant of GPT-5.3-Codex designed specifically for real-time interactive coding. This marks a significant milestone: it is OpenAI's first production deployment on hardware other than Nvidia, running on Cerebras' Wafer Scale Engine 3.

The release signals a strategic shift in how AI-powered coding tools operate. Rather than prioritizing raw capability at the expense of latency, Codex-Spark focuses on making AI assistance feel instantaneous, enabling developers to integrate AI into their edit-compile-debug loops without waiting.

Key Technical Specifications

GPT-5.3-Codex-Spark delivers over 1,000 tokens per second when running on the Cerebras WSE-3 chip, a massive silicon wafer containing more than 4 trillion transistors. The model features a 128k token context window and operates as a text-only model at launch.

OpenAI achieved substantial infrastructure improvements alongside the model release. Through the introduction of persistent WebSocket connections and targeted optimizations, the company reduced overhead per client/server roundtrip by 80%, per-token overhead by 30%, and time-to-first-token by 50%.

Benchmark Performance

Despite being a smaller model optimized for speed, Codex-Spark demonstrates strong coding capabilities. On Terminal-Bench 2.0, it achieved 77.3% accuracy, a significant improvement over GPT-5.2-Codex's 64% score. The model also surpasses the reasoning capabilities of both GPT-5.2-Codex and the standard GPT-5.2 model in coding-specific tasks.

On SWE-Bench Pro, the model shows competitive performance while completing tasks in a fraction of the time compared to the full GPT-5.3-Codex. This speed-accuracy tradeoff positions it ideally for interactive development scenarios.

Beyond Basic Code Completion

Codex-Spark is not limited to simple code generation. The model handles a broad range of development tasks including debugging, deploying, monitoring, writing product requirement documents, editing copy, conducting user research, writing tests, and tracking metrics. This expanded scope makes it a comprehensive development companion rather than just a code autocomplete tool.

The Cerebras Partnership

The deployment on Cerebras hardware represents OpenAI's first move away from exclusive reliance on Nvidia GPUs for production workloads. The partnership, announced in January 2026, leverages Cerebras' purpose-built Wafer Scale Engine 3 accelerator, which is specifically designed for low-latency inference workloads.

This diversification of hardware partnerships could have significant implications for the AI infrastructure landscape, potentially reducing dependency on a single chip manufacturer and opening new performance optimization pathways.

Availability and Access

Codex-Spark is currently available as a research preview exclusively for ChatGPT Pro subscribers. Usage during this preview phase does not count toward standard API limits, with separate rate limits that may be adjusted. API access is expected to arrive soon, at the same pricing as the standard GPT-5.3-Codex model.

What This Means for Developers

The release of Codex-Spark represents a paradigm shift in AI-assisted development. With sub-second response times and strong coding accuracy, developers can now use AI assistance in real-time editing workflows where even a few seconds of latency would break concentration. The 128k context window ensures the model can understand large codebases, while the speed optimizations make it practical for continuous use throughout a coding session.

For teams evaluating AI coding tools, Codex-Spark offers a compelling combination of speed and capability that addresses one of the most common complaints about AI coding assistants: they are powerful but too slow for interactive use.