DeepSeek V4 Flash API Cost: Thinking Mode Corrupts Strict JSON

DeepSeek released three versions of its V4 model: Flash 0731, Flash Preview, and V4 Pro. All three share the same tokenizer, cache, and thinking machinery, but they differ in pricing and failure modes. The Flash 0731 build is the cheapest at $0.14 per million input tokens and $0.28 per million output, while the V4 Pro costs $0.435 per million input tokens and $0.87 per million output.

A critical issue was discovered with the Flash 0731 build when using thinking mode and strict JSON schema. In 8 out of 13 runs across two request paths, integer fields were corrupted. Turning off thinking fixed the issue, reducing token usage to a seventh of the default. The retrained 0731 build sharpened the off-switch cliff, causing the 2-hop math to fall from 6/6 to 0/6.

The cache serves 1,024-token pages from a 1.1K-token floor, with hits occurring approximately 0.3 seconds after priming. Entries remain valid for up to 45 minutes. Disabling thinking entirely (enable_thinking: false) resolved the corruption issue, executing the same task in about 44 completion tokens compared to the default's 328.

The preview build performed flawlessly with thinking on, while V4 Pro showed corruption in 2 out of 4 runs. The Flash 0731 build was the hardest hit, with the corruption issue being the most severe. The problem is not related to the thinking budget but to the presence of thinking itself.

In summary, the three V4 builds share the same underlying infrastructure but differ in pricing and failure modes. The Flash 0731 build, when using thinking mode with strict JSON schema, experiences JSON corruption. Disabling thinking resolves the issue, making thinking and strict structured output mutually exclusive for this model. The lesson is to treat thinking and strict structured output as incompatible options when using DeepSeek V4 Flash.