Introducing GPT-5.3-Codex-Spark

• GPT-5.3-Codex-Spark: smaller Codex built for real-time coding.

  • 1,000 tokens/sec on Cerebras Wafer-Scale Engine 3.

  • Research preview Feb 2026; Pro subscribers via Codex app, CLI, VS Code.
• First milestone of OpenAI × Cerebras partnership (announced Jan 2026).
  • WSE-3: single 125-petaflop wafer-scale chip purpose-built for inference.
• Outperforms GPT-5.1-Codex-mini on SWE-Bench Pro and Terminal-Bench 2.0.
  • Completes same tasks in fraction of time vs GPT-5.3-Codex.
• Specializes in precise edits, plan revisions, contextual code questions.
  • Near-instant UI iteration: layouts, styling, live testing feedback.
• Two Codex modes planned: long-horizon reasoning + real-time collaboration.
• API rolling to select design partners; Cerebras targeting frontier models 2026.

X discourse

• @AriX: “First time I’ve seen an LLM operate a GUI as fast as a person, and it’s surreal.” (5117 likes)
• @anishmoonka: “Karpathy: Shift to agents happened Dec 2025. Humans now bottleneck; unit of work is a whole feature.” (872 likes)
• @chatgpt21: “GPT 5.4 w/codex made sizable dent in automating tedious labor. 20% better will be extremely disruptive.” (488 likes)
• @petergyang: “Live demo: Building in seconds with Spark. How the Codex team operates without traditional specs.” (455 likes)
• @cerebras: “What can you build with gpt-5.3-codex-spark? @jxnlco from @OpenAI demos 3 real workflows.” (217 likes)
• @haider1: “Codex 5.3 pushed coding agents to new level; curve feels much steeper now.” (210 likes)

· ** · Read on openai.com