SAN FRANCISCO — February 13, 2026 — OpenAI has introduced GPT-5.3-Codex-Spark, a major new iteration of its Codex family designed specifically for real-time coding collaboration and rapid edits. The research preview makes this smaller, latency-optimized version of GPT-5.3-Codex available to developers while OpenAI and partner Cerebras continue building out ultra-low-latency infrastructure. 

GPT-5.3-Codex-Spark is engineered to deliver near-instant responses, processing more than 1,000 tokens per second on purpose-built Cerebras hardware for high-speed inference. The model maintains strong coding capabilities while enabling interactive workflows where users can redirect or pause execution and see results immediately — a behavior tailored to live, IDE-style collaboration. 

Codex-Spark supports a large 128k context window and is currently text-only, reflecting its focus on fast code authoring, logic refactoring, and interface updates rather than broader multimodal tasks. During the research preview, usage is governed by a dedicated rate limit that may adjust based on demand and reliability requirements. 

OpenAI said Codex-Spark complements existing Codex models by supporting both real-time editing and longer-running development tasks — allowing developers to choose the right performance profile for different coding scenarios. This milestone also marks the first product to emerge from OpenAI’s partnership with Cerebras, which provides a latency-first serving tier using the Wafer Scale Engine. 

GPT-5.3-Codex-Spark is currently available as a research preview to ChatGPT Pro users in the latest Codex app, CLI, and VS Code extension, with API access for select partners as OpenAI expands capacity and hardens the experience. 

OpenAI said this model is a step toward making AI-assisted programming feel like a true real-time collaboration, offering developers an instantaneous feedback loop traditionally found in human teamwork but seldom seen in AI workflows. 

Source: OpenAI