OpenAI announced GPT-5.3-Codex-Spark on 12 February 2026, describing it as a specialized coding model with enhanced performance on software engineering tasks [source]. The changelog states the model was trained on "proprietary internal repositories" and benchmarked against previous Codex iterations, showing improvements in code completion and debugging accuracy.

No independent verification of the claimed benchmarks has been published. The announcement does not specify which repositories were used for training, whether they included third-party code, or what licensing terms apply to outputs generated from potentially copyrighted material.

The model is available via API to enterprise customers. OpenAI's documentation states that outputs "may reflect patterns from training data" but does not detail what disclosure mechanisms exist for users to determine if generated code matches existing copyrighted works.

This follows a pattern observed across multiple providers in late 2025 and early 2026, where coding-focused models have been trained on increasingly opaque datasets. In December 2025, GitHub Copilot faced scrutiny over code suggestions that closely matched GPL-licensed repositories without attribution. Anthropic's Claude 3.7 Code, released in January 2026, similarly declined to specify training sources beyond "publicly available and licensed code."

The Codex-Spark announcement includes no information on drift monitoring, version stability, or how the model handles requests to reproduce training data. OpenAI has not responded to questions about whether the model was tested for memorization of proprietary code or what safeguards prevent verbatim reproduction of licensed material.

The model is designated severity-3 due to transparency gaps in training data provenance and potential intellectual property implications for enterprise users.