OpenAI Launches GPT-5.3-Codex-Spark
OpenAI has launched GPT-5.3-Codex-Spark, a model engineered for real-time code generation and agent orchestration that leverages Cerebras’ WSE-3 chips for up to 15x faster inference. The company also retired GPT-4o, making GPT-5.2 the new default for ChatGPT. Technical documents suggest the new model manages parts of its own training pipeline, hinting at early forms of recursive self-improvement.
- The partnership with Cerebras is part of a multi-year, 750-megawatt agreement to build out an alternative to GPU-based infrastructure for low-latency workloads. This move signals OpenAI's strategy to use specialized hardware for specific tasks, aiming to reduce inference bottlenecks. - Cerebras’ WSE-3 chip is a single wafer-scale processor with 4 trillion transistors and 900,000 AI-optimized cores. It features 44GB of on-chip SRAM, providing 21 PB/s of memory bandwidth, designed to keep all model parameters on-chip to minimize latency. - The "Spark" model is a smaller, distilled version of the full GPT-5.3-Codex, optimized for generating over 1000 tokens per second. This speed comes with a trade-off in reasoning, as it scores significantly lower on complex benchmarks like SWE-Bench Pro and is intended for interactive tasks like boilerplate generation, not system architecture. - Access to GPT-5.3-Codex-Spark is currently limited to a research preview for ChatGPT Pro users and is not yet available via the commercial API. It operates under a separate rate limit, indicating that the specialized Cerebras hardware is a distinct capacity pool from OpenAI's primary GPU clusters. - The concept of recursive self-improvement (RSI) is an active area of research where a model attempts to rewrite its own code to enhance its capabilities. However, some research indicates limitations due to "entropic drift," where models cannot bootstrap to greater intelligence without new external data or feedback. - The retirement of
