OpenAI launches GPT-5.3-Codex-Spark, a lightweight version of flagship coding model

On February 12, 2026, OpenAI announced GPT-5.3-Codex-Spark, a lightning-fast, lightweight version of its flagship coding model. It is designed for instant interaction; this research preview marks the first tangible result of OpenAI’s new partnership with Cerebras Systems.

Unlike traditional frontier models that excel at long, complex tasks but feel sluggish in live sessions, Codex-Spark delivers over 1,000 tokens per second. Developers can now interrupt the AI mid-generation, redirect it, and see targeted code edits appear almost instantly. The model avoids unnecessary full rewrites or automatic test runs unless specifically asked, making it ideal for rapid iteration, debugging, and collaborative coding.

The speed gains come from deep optimizations across the entire stack. It brings an 80% reduction in client/server round-trip overhead, 30% lower per-token latency, and 50% faster time-to-first-token.

Persistent WebSocket connections are now the default, and the entire inference pipeline has been streamlined. These improvements will eventually benefit all OpenAI models, not just the new Spark variant.

Codex-Spark runs on Cerebras’ Wafer Scale Engine 3, the massive chip built specifically for extreme-speed inference. It complements OpenAI’s GPU clusters rather than replacing them, creating a hybrid system where fast models handle real-time work while larger models tackle heavy reasoning in the background.

The model scored strongly on SWE-Bench Pro and Terminal-Bench 2.0, showing it can handle real agentic software engineering tasks at a fraction of the usual latency.

It ships with a 128k context window (text-only in this preview) and includes the same safety training as OpenAI’s mainline models, such as cyber-specific safeguards. On the other hand, Independent evaluations confirmed it does not reach high-risk thresholds in cybersecurity or biology.

Initially, ChatGPT Pro users can try GPT-5.3-Codex-Spark in the latest Codex app, CLI, and VS Code extension. Meanwhile, a small group of API design partners also has access. Furthermore, Usage comes with separate rate limits and may involve brief queuing during peak times.

(source)