GPT-5.3-Codex-Spark

GPT-5.3-Codex-Spark is a research preview of OpenAI's ultra-fast coding model announced on 12 February 2026. It is a smaller variant of the flagship GPT-5.3-Codex and OpenAI's first model specifically designed for real-time work with the Codex coding agent. The model is served on the Cerebras Wafer-Scale Engine 3 accelerator, allowing it to generate over 1000 tokens per second — a fundamental change in how AI model interaction feels.

Technical profile

Codex-Spark has a 128k context window and is text-only at launch. It is tuned for making precise, minimal code edits — it does not automatically run tests unless asked. The model is steerable in real time — you can interrupt it, redirect it, ask questions and see results almost instantly.

Benchmarks

According to OpenAI, Codex-Spark scores stronger than GPT-5.1-Codex-mini on SWE-Bench Pro and Terminal-Bench 2.0 while finishing the same tasks in a fraction of the time required by GPT-5.3-Codex. Codex-Spark is the first member of an ultra-fast model family from OpenAI.

Availability

Codex-Spark is available as a research preview for ChatGPT Pro users in the Codex app, CLI and VS Code extension. API access is rolling out to a small set of design partners. The model runs on a separate ultra-low-latency path (Cerebras) with its own rate limit that does not count against standard Codex limits.

Partnership with Cerebras

Codex-Spark is the first outcome of OpenAI's partnership with Cerebras Systems, announced in January 2026. Sean Lie, CTO and co-founder of Cerebras, described the collaboration as an attempt to discover what new interaction patterns, use cases and fundamentally different model experiences become possible with fast inference. GPUs remain foundational to OpenAI's training and the bulk of its inference — Cerebras complements that architecture with a low-latency tier for workflows where responsiveness matters as much as intelligence.

Latency optimizations

Together with the Codex-Spark release, OpenAI rolled out end-to-end latency improvements across the entire Codex serving stack — a persistent WebSocket connection, optimizations inside the Responses API and a rewritten response-streaming path. The combined effect: −80% per-roundtrip overhead, −30% per-token overhead and −50% time-to-first-token. These improvements will gradually be made the default for all Codex models.