DeepSeek V4 Flash has a defect in its 0731 build where integer fields in structured JSON output can be corrupted when thinking is enabled. This issue occurs specifically with strict JSON schemas and was observed in a majority of test runs. Disabling thinking, however, resolves the corruption and significantly reduces token usage. The 0731 retrain also introduced a sharper performance drop for multi-step tasks when thinking is off. The model features a cache that stores 1,024-token pages, with hits occurring quickly and entries persisting for at least 45 minutes. Pricing remains consistent at $0.14 per million input tokens and $0.28 per million output tokens, with cache hits being significantly cheaper. Compared to other V4 builds, the 0731 version is most affected by the structured output corruption. The effort dial on the API did not yield measurable differences in performance or token usage. Two-hop math, a complex task, now heavily relies on thinking being enabled for the 0731 build. The context window is substantial, accommodating large prompts, and output limits appear to be negotiable. Users requiring reliable structured output should disable thinking on the 0731 build until the defect is addressed.