DeepSeek's budget AI model outperformed its more expensive flagship on a demanding coding challenge, delivering an 1.83x speed improvement versus 1.06x in new benchmark tests published by The New Stack. The finding raises questions about why the Chinese lab's flagship V4-Pro model costs three times more than V4-Flash when the cheaper version demonstrated superior reasoning on complex tasks. Both models tied on simpler assignments, but diverged sharply when asked to optimize performance in real production code.

The tests ran both models through three professional coding jobs using Rich, a Python terminal-formatting library, inside the OpenCode terminal agent. On the first task, a bug fix for a macOS import crash, Flash completed the work in 59 seconds using 119,900 tokens across 6 API requests for under a cent, while Pro took 1 minute 1 second, made 13 requests, and consumed 159,200 tokens for about a cent. Both delivered identical one-line fixes. On the second task, building a new feature to strip leading whitespace while preserving text formatting, Flash finished in 1 minute 29 seconds on 191,200 tokens and 6 requests for under a cent, while Pro required 1 minute 43 seconds, 333,600 tokens, and 23 requests for roughly two cents. Both implementations passed all eight edge cases tested. The performance optimization test separated them dramatically: Flash worked for 27 minutes, burned 15.4 million tokens across 128 API requests, and cost about eight cents, while Pro ran 15 minutes 21 seconds, used 4.7 million tokens over 57 requests, and cost seven cents. Flash's total token consumption across all three tests was triple Pro's, yet final bills landed at nine cents versus ten cents because Flash's per-token rates run one-third of Pro's.

The benchmark author verified Flash's claimed 1.84x speedup independently at 1.83x, confirming the performance gain was real. Flash built its own testing framework to compare old and new output across 37 table types, caught and fixed one of its own bugs during development, and cleaned up 13 type-checker errors, though it incorrectly claimed its new output was byte-identical to the original when one edge case showed invisible color codes written in different order. Pro's 1.06x measured improvement matched its 3.3% reported gain, and its output truly was byte-identical including the case Flash missed, but Pro overlooked the larger optimization opportunity Flash discovered in the same code. According to the report, "the cheaper model got creative with this one, and the expensive model kept the work more on the surface."

The results suggest better reasoning capability correlates with higher token consumption, which could reshape AI spending as models tackle harder problems. Flash achieved superior performance on the reasoning-heavy task precisely because it consumed three times the tokens, running hundreds of tool calls and building elaborate testing infrastructure where Pro took a more direct path. The report notes DeepSeek currently displays a banner warning that prices will increase significantly soon, raising the possibility that Flash's low pricing aims to establish user habits around its stronger reasoning before rate hikes arrive. The parent company sets pricing as strategy rather than cost reflection, meaning current rate differences reveal commercial positioning more than underlying expense. Companies leaning on AI coding agents for complex optimization work may face sharply higher bills if reasoning quality scales with token burn, regardless of which vendor they choose.