GPT-5.3-Codex-Spark Release Notes

OpenAI has released a research preview of GPT-5.3-Codex-Spark, a small-scale version of GPT-5.3-Codex designed specifically for real-time, interactive coding. By leveraging a partnership with Cerebras, the model delivers over 1,000 tokens per second, enabling developers to perform targeted edits and reshape logic with near-instant feedback.

Real-Time Coding Capabilities

GPT-5.3-Codex-Spark is optimized for interactive workflows where latency is as critical as intelligence. Unlike larger models designed for autonomous, long-running tasks, Codex-Spark is tuned for a lightweight working style, focusing on minimal, targeted edits and avoiding automatic test execution unless explicitly requested. This allows developers to interrupt or redirect the model in real time to rapidly iterate on interfaces and logic.

Performance and Benchmarks

Despite its smaller size, Codex-Spark maintains high capability for real-world software engineering tasks. On the SWE-Bench Pro and Terminal-Bench 2.0 benchmarks, which evaluate agentic software engineering, the model demonstrates strong performance while completing tasks in a fraction of the time required by the full GPT-5.3-Codex model.

Infrastructure and Latency Optimizations

To support real-time collaboration, OpenAI implemented end-to-end latency improvements across its entire request-response pipeline. These optimizations benefit all models and include:

  • Pipeline Streamlining: Rewritten inference stack and optimized session initialization to reduce time-to-first-token by 50%.
  • Network Efficiency: The introduction of a persistent WebSocket connection and optimizations to the Responses API, which reduced per-client/server roundtrip overhead by 80% and per-token overhead by 30%.

Hardware Integration with Cerebras

GPT-5.3-Codex-Spark is powered by the Cerebras Wafer Scale Engine 3, a purpose-built AI accelerator designed for high-speed inference. While GPUs remain the primary foundation for cost-effective broad usage, Cerebras hardware provides a latency-first serving tier that complements GPUs to tighten the iteration loop for developers.

"What excites us most about GPT-5.3-Codex-Spark is partnering with OpenAI and the developer community to discover what fast inference makes possible—new interaction patterns, new use cases, and a fundamentally different model experience."

— Sean Lie, CTO and Co-Founder of Cerebras

Availability and Technical Specifications

Codex-Spark is currently available as a research preview for ChatGPT Pro users via the Codex app, CLI, and VS Code extension. It is also available in the API for a small group of design partners.

Technical Specifications:

  • Context Window: 128k tokens.
  • Modality: Text-only.
  • Safety: The model underwent standard deployment evaluations and does not meet the Preparedness Framework threshold for high capability in cybersecurity or biology.

Future Roadmap

OpenAI intends for Codex-Spark to be the first of a family of ultra-fast models. The long-term goal is a hybrid Codex experience that blends long-horizon reasoning and execution with real-time collaboration. This will allow users to maintain an interactive loop while delegating complex, background work to sub-agents or fanning out tasks to multiple models in parallel.

Sources