OpenAI Coding Model Runs 15x Faster on New Hardware

Waiting for AI to generate code might soon feel like a distant memory. OpenAI released GPT-5.3-Codex-Spark on Thursday, a specialized coding model that delivers more than 1,000 tokens per second, leaving older versions in the dust.

The new model runs on dinner plate-sized chips from Cerebras, marking OpenAI's first production release on non-Nvidia hardware. For developers building apps and websites, this speed boost means less thumb-twiddling and more creating.

ChatGPT Pro subscribers can access Codex-Spark now through the Codex app and VS Code extension for $200 per month. OpenAI designed it specifically for coding tasks, trading the broad knowledge of larger models for pure speed.

The company tested Spark on two software engineering benchmarks, where it reportedly outperformed older models while finishing tasks much faster. When Ars Technica tested coding agents in December, speed became the deciding factor between tools that felt helpful and ones that felt frustrating.

OpenAI's fastest previous models on Nvidia chips topped out around 147 to 167 tokens per second. Codex-Spark's leap to over 1,000 represents a dramatic improvement, though Cerebras has clocked even higher speeds with other models, suggesting room to grow.

The Ripple Effect

This release signals more than just faster code generation. OpenAI has been quietly building relationships with AMD, Amazon, and now Cerebras throughout the past year, reducing its reliance on Nvidia's dominant position in AI chips.

The company even started designing its own custom chips for future production. These moves reflect a broader shift in the AI industry, where competition is pushing companies to explore new hardware partnerships and drive innovation beyond a single supplier.

For the developers who spend hours inside code editors, AI tools have transformed from curiosities into daily teammates. Speed determines whether these assistants enhance creativity or interrupt flow, making latency the new battleground among OpenAI, Google, and Anthropic.

The rapid iteration shows how seriously these companies take the coding assistant race. OpenAI released GPT-5.2 in December after CEO Sam Altman issued an internal warning about competitive pressure, then shipped GPT-5.3-Codex days ago, and now Spark arrives with its speed focus.

The hardware diversity matters beyond corporate strategy too. When AI companies spread across different chip makers, it creates more paths for innovation and potentially more stable supply chains for the tools millions of developers increasingly depend on.

Codex-Spark shows what happens when software and hardware teams solve problems together, turning a speed bottleneck into a competitive advantage that makes creating software faster and more joyful for everyone involved.