TL;DR
OpenAI has released GPT-5.3-Codex-Spark, its first model built specifically for real-time coding. Running on Cerebras’ Wafer Scale Engine 3, it delivers over 1,000 tokens per second — fast enough for developers to collaborate with the model interactively, redirecting it mid-task.
Speed Meets Capability
Codex-Spark is a smaller version of GPT-5.3-Codex, optimised for low-latency inference. The model is designed to complement OpenAI’s existing long-running coding agents with a fast, interactive mode suited to targeted edits, logic reshaping, and interface refinement.
“With Codex-Spark, Codex now supports both long-running, ambitious tasks and getting work done in the moment,” OpenAI said.
The model features a 128k context window (text-only at launch) and is available as a research preview for ChatGPT Pro users through the Codex app, CLI, and VS Code extension.
Infrastructure Overhaul
Beyond the model itself, OpenAI reworked its inference stack to reduce latency across all models. Key improvements include:
- 80% reduction in overhead per client/server round-trip
- 30% reduction in per-token overhead
- 50% reduction in time-to-first-token
These gains came from introducing persistent WebSocket connections and rewriting how sessions initialise. The WebSocket path is enabled by default for Codex-Spark and will become the standard for all models.
The Cerebras partnership, announced in January, provides the low-latency serving tier. GPUs remain the backbone for training and cost-effective broad inference, with Cerebras handling workflows where speed is the priority.
Looking Forward
OpenAI’s roadmap points toward blending real-time and long-running modes: Codex will keep developers in a tight interactive loop while delegating background work to sub-agents. For UK developers evaluating coding AI tools, Codex-Spark competes directly with Anthropic’s Claude Code (highlighted by Spotify this week) and signals that inference speed is becoming a key differentiator alongside raw capability.