V4-Flash vs. V4-Pro: DeepSeek promised better and cheaper. It’s true, but not how I expected.
The New Stack Jessica Wachtel ● Covered by 4 sources
DeepSeek’s cheaper V4-Flash beat or matched V4-Pro in coding tests. The surprise: the pricey model only mattered once, and even then cost barely separated them.
Based on reporting by The New Stack, Jessica Wachtel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek refreshed V4-Flash and moved it out of preview, with docs claiming it beat V4-Pro preview on coding and agentic benchmarks. That sounds odd on its face because Flash costs about a third as much: $0.14 per million input tokens and $0.28 per million output, versus $0.435 and $0.87 for Pro.
So the comparison was simple enough to be useful. Both models were dropped into OpenCode, the open source terminal coding agent, and given the same three jobs against the Rich Python library: a bug fix, a feature build, and a performance optimization. The setup was kept tight: separate repo clones, fresh sessions, identical prompts, and tracking for time, token use, API requests, and cost.
On the bug fix, both models found the same issue. Rich was crashing on macOS because it called os.getcwd() during import and only caught FileNotFoundError, while File Provider folders like Dropbox and iCloud can raise PermissionError instead. Both models widened the exception handler to OSError, wrote regression tests, and got all 957 tests passing. Flash was a little cleaner on implementation detail, but the end result was the same. Call it a tie.
The feature work was stranger. Adding lstrip() and strip() to Rich’s Text class meant preserving styles while text shifts left, which is where models can easily scramble formatting. Both got the job done and passed eight edge cases. But Pro spent a lot more effort getting there: more requests, more tokens, and a subagent that made 13 tool calls before the main work even started. Flash did the same job with less ceremony and less cost, so it won on efficiency.
The performance task is where the gap opened. Flash spent 27 minutes and built its own testing rig, compared old and new output across 37 table types, fixed its own mistakes, and reported a 1.84x speedup; the measured result came out at 1.83x. Pro spent 15 minutes 21 seconds, found a smaller optimization, and reported 3.3%; the measured speedup was 1.06x. Pro’s output was byte-identical. Flash’s claim of identical output was not quite true, because one color-terminal case reordered invisible codes, even though the screen looked the same.
The final bill is the weird part. Flash used about 15.7 million tokens across the whole run and cost $0.09. Pro used about 5.2 million and cost $0.10. Cheap rates plus heavy token use cancel out. DeepSeek’s bargain model is the one that gets creative when the job gets hard, and that’s the real story here.
My take — AI-written commentary, not fact-checked reporting
This is the old AI pricing trick in cleaner clothes: sell the cheaper model as the obvious choice, then let the harder work reveal why the expensive one still exists. Flash looks like the bargain until the reasoning task shows up and the token meter starts spinning like a bad taxi. The industry keeps pretending cost and capability are neatly separable; they aren’t, and the bill always arrives with a smile.
Read more about this at: The New Stack
