GPT-5.3-Codex-Spark

The model is explicitly identified as a 'streamlined' and 'smaller' variant of the GPT-5.3-Codex family, developed in collaboration with Cerebras Systems. While the base model lineage is clear, specific architectural modifications beyond being a 'dense transformer' are not fully detailed. Documentation mentions it was 'physically smaller/pruned' to fit the SRAM requirements of the Cerebras WSE-3, but the exact pruning or distillation methodology remains proprietary and lacks deep technical disclosure.

OpenAI provides almost no specific information regarding the training data for the Spark variant. It is described as having the 'same safety training' as mainline models, implying a shared data heritage with GPT-5.3-Codex, but there is no public breakdown of data sources, proportions (e.g., code vs. text), or specific filtering methodologies used for this specialized coding variant. Claims of 'high-quality' or 'cyber-relevant' training are unverifiable marketing assertions.

While the model is accessible via the Codex app and CLI, there is no public documentation regarding its specific tokenizer vocabulary size or training data alignment. It is stated to be 'text-only' at launch, but technical specifications for the tokenizer (such as whether it uses the standard tiktoken/o200k_base or a specialized coding-optimized version) are not disclosed.

The parameter count is officially 'Unknown'. Documentation vaguely describes it as a 'smaller version' and 'physically smaller' to accommodate the Cerebras WSE-3 hardware, but no specific numbers for total or active parameters are provided. This lack of transparency makes it impossible to verify the model's actual density or efficiency claims.

The hardware used for inference (Cerebras WSE-3) is heavily documented, but the actual training compute metrics—such as GPU/TPU hours, training duration, or carbon footprint—are entirely absent. While the partnership with Cerebras for deployment is a core marketing pillar, the upstream training resources remain undisclosed.

OpenAI cites scores on SWE-Bench Pro (58.4%) and Terminal-Bench 2.0 (77.3%), but evaluation code and exact prompts are not public. There is significant confusion in secondary reporting, with some sources claiming it matches the full GPT-5.3-Codex while others (and OpenAI's own blog) state it underperforms. The lack of a clear reproduction path or third-party verification for these specific 'Spark' results justifies a low score.

The model consistently identifies as part of the GPT-5.3-Codex family and is transparent about its role as a 'latency-first' variant. It acknowledges its limitations compared to the larger model, specifically noting it is better for 'targeted edits' than 'autonomous code rewrites.' However, it lacks deep version-specific self-awareness in its output beyond the family name.

The model is governed by a 'Proprietary' license with no open-weights or open-source availability. Terms are tied to the ChatGPT Pro subscription ($200/mo) and a 'latency-first' serving tier. There is no public license document for the weights or architecture, and commercial use is restricted to OpenAI's platform terms.

Hardware requirements for inference are well-documented in the context of the Cerebras WSE-3 partnership, including the 128k context window and 1,000+ tokens/sec throughput. However, because the model is not available for local deployment, VRAM requirements for standard hardware (FP16/Q4) are irrelevant/undisclosed, and quantization tradeoffs are handled entirely server-side without user visibility.

The model is currently in 'research preview' with no public changelog or semantic versioning beyond the 'Spark' designation. OpenAI has already made silent performance adjustments (claiming a 30% speed increase shortly after launch), which indicates a high potential for unannounced behavior drift without a formal tracking mechanism for users.