DeepSeek V4 Flash vs Pro on Practical Prompt Tests
DeepSeek V4 Pro narrowly beat DeepSeek V4 Flash in this run of model-prompt-tests: 4.50 average peer-judge score to Flash’s 4.38 across 13 practical prompts.
The score gap is small. The candidate-answer cost gap is not. Pro recorded $0.2148 versus Flash’s $0.0100, roughly 21 times as much for a +0.12 score difference.
That makes Flash the sensible default for this prompt set. Pro has a case when the extra edge in writing and instruction precision matters, but it did not create a clear advantage on the core engineering prompts.
This is one same-lineage comparison, not a general model ranking. Both models used the same OpenRouter path, the same 13 repository-local prompts, and the same two external judges. You can read how the harness works in Benchmarking Model Prompt Tests.
The full run is available in the DeepSeek V4 Flash vs Pro artifact directory: run config, summary JSON, per-prompt results, and raw outputs with score rationales.
How I ran it
The batch was 2026-08-13T21-13-26-015Z-85hr0k. It compared these two candidates through OpenRouter:
openrouter:deepseek-v4-flash, mapped todeepseek/deepseek-v4-flash-0731openrouter:deepseek-v4-pro, mapped todeepseek/deepseek-v4-pro-0813
Each candidate answered 13 prompts covering architecture, code review, debugging, safety, planning, writing, data analysis, and instruction following. GPT-5.5 and Claude Sonnet 5 then scored the outputs against the prompt rubrics. Neither candidate was a judge, so this run does not have a self-judging problem.
The headline score is the mean of each prompt’s median peer-judge score. All 26 candidate responses completed, and all 52 judge rows returned valid scores.
I raised the candidate token cap to 32768 after an earlier 4096-token attempt spent its budget on hidden reasoning and returned empty content. That is a harness detail, not a refusal from either model. Pro also required peer ranking to fall back from JSON Schema to a plain-JSON contract.
Headline results
| Model | Completed runs | Avg score | Avg latency | Avg output tokens | Total cost |
|---|---|---|---|---|---|
openrouter:deepseek-v4-pro |
13 | 4.50 | 2.93 s | 4,954 | $0.2148 |
openrouter:deepseek-v4-flash |
13 | 4.38 | 2.83 s | 3,505 | $0.0100 |
The latency difference was effectively a tie in this setup. Pro was a little slower and substantially more verbose. Cost is recorded OpenRouter usage for this batch, not a promise about another provider, model snapshot, or future price.
Where the result was close
Both models scored a perfect 5.00 on four core engineering prompts:
architecture/database-choicecode-review/senior-pr-reviewdebugging/javascript-debouncesafety-risk/failed-production-migration
For schema choice, PR review, a concrete JavaScript bug, and a risky migration, the judges could not distinguish them. That is why the score gap needs to stay in proportion. I would not pay Pro’s price for those tasks on the evidence from this run.
Pro earned its narrow lead through five prompts: an internal announcement, a San Francisco trip plan, constrained app names, a five-bullet summary, and a six-hour SQL lesson plan. It looked better at writing, personalisation, format-heavy summarisation, and keeping a teaching plan realistic.
Flash won three technical or analytical prompts: the Pomodoro build, a business-metrics diagnosis, and an evaluation-rubric design. The judges preferred its complete runnable timer, its focus on churn and support-ticket growth, and its slightly cleaner rubric anchors.
The shared weak spot
Both models scored 2.00 on the ambiguous team-notes-app request. The rubric rewarded clarifying questions, explicit assumptions, and a scoped MVP. Both answers jumped to a large full-stack implementation instead.
That is more useful than treating the comparison as a winner-takes-all score. If I want clarification before a build, neither model did it reliably here. The system prompt or the first turn needs to require it.
Both also lost points on the six-hour SQL-plan prompt because the proposed lesson plans did not fit the time well. More content was not the answer in either case.
Peer ranks are not the headline
The harness also ran anonymised peer ranking after scoring. Each candidate ranked the two answers with brand labels hidden. The final Borda totals were tied: Flash 39, Pro 39.
That number stays a side signal. With two candidates, each ranker usually preferred its own output. Peer ranks are useful for finding prompts where rankings and rubric scores agree or disagree, but they are relative to this roster and should not replace the rubric-based avgScore.
There were useful agreements. Both rankers preferred Flash’s business-metrics diagnosis, and both preferred Pro’s SQL plan and internal announcement. There were also mismatches, including a ranked Flash win on the debounce task even though rubric scores tied at 5.00. That is exactly the kind of disagreement worth inspecting in the artifacts, not averaging into a new headline metric.
The run also produced four optional chairman-synthesis answers. Those are combined answers for readers, not a third scoring stage, so they are not part of this comparison.
What I would choose
For this prompt set, I would start with Flash. Its score was within a tenth of Pro’s, it tied Pro on the core engineering prompts, and its candidate-answer cost was much lower.
Pro is the option when the task is closer to polished writing, a tailored travel plan, careful formatting, or a constrained teaching plan. Even there, this is a small one-batch result, not proof that the pattern will hold.
The next useful run is repetition. Three or more batches would show whether the +0.12 gap is stable. I would also add a third non-DeepSeek ranker and use human review for the five-bullet prompt, where the two judges differed by two points on Pro’s answer.
For now the narrow conclusion is enough: Pro scored slightly higher in this August 13, 2026 batch. Flash delivered nearly the same measured quality for a fraction of the candidate cost.