DeepSeek V4 Pro vs V4.1 Flash: Same Answers, 5 to 9 Times the Cost

What we ran

On October 28 and 29, 2026 we ran DeepSeek V4.1 Flash (API model deepseek-flash) and DeepSeek V4 Pro (deepseek-v4-pro) on the same 720 tasks each. Every answer was checked by plain Python code, not by another LLM. A failed answer got up to two retries, and every attempt was billed, so the cost per correct answer includes the misses.

The tasks come from our open LLM cost per task benchmark. Nine everyday tasks with 50 items each: invoice field extraction, strict JSON, text-to-SQL, intent labels, bug fixes, tool calls, questions over a 30K-token handbook, record cleanup and regex writing. Then six harder versions of those, 270 items in all. The receipt-image task was skipped because neither model takes images. We used provider defaults: no temperature, no effort setting.

Costs below use DeepSeek's Peak list prices from its pricing page, the same rates our DeepSeek API pricing page shows:

Same accuracy, 4.9x to 9.2x the cost

Both models reached 100% within three attempts on both tiers. First-try success was 95% to 100% on every task for both, never more than one item apart. So on these tasks you are not buying accuracy with V4 Pro. You are buying the same answers at a higher price.

In absolute terms the whole run is cheap: $0.32 at list price for Flash's 720 items and $2.24 for V4 Pro's. The ratio is what matters once volume grows. At 100,000 everyday tasks a month, Flash comes to about $35 and V4 Pro to about $170. On hard tasks it is about $60 against $548.

The gap also grows with difficulty. On everyday tasks V4 Pro cost 4.9x as much per correct answer. On the hard tier it was 9.2x, and on the hardest single task, regex writing with leap years, Roman numerals and lookaheads, it was 14x.

Where V4 Pro's money goes: reasoning tokens

If both models wrote the same number of tokens, V4 Pro would cost between 3.3x and 7.3x as much, depending on how much of the bill is output, fresh input or cached input. It cost more than that because it writes far more output, and almost all of the extra is reasoning. You never see those tokens in the reply, but they bill at the output rate.

On the hard tier V4 Pro's mean output per attempt was 1.3x to 4.3x Flash's. On hard regex items it averaged 4,294 output tokens per attempt against Flash's 1,003. A 3.3x output price times a 4.3x token count is roughly 14x, which is the gap we measured on that task.

cost ratio per correct answer ≈ price ratio × token ratio ÷ success-rate ratio

hard regex task:  3.3 × 4.3 ÷ ~1.0  ≈  14x

This is the part a price table can't show you. Two models can be 3.3x apart on the rate card and 9x apart on the invoice, and the only way to find the token ratio is to run your own prompts and read the usage field on every response.

Empty answers still cost money

Our harness caps output at 16,000 tokens per call. Three V4 Pro attempts (two hard regex items and one bug fix) hit that cap while still reasoning and returned nothing. Each of those calls was billed for 16,000 output tokens, about $0.063 at Peak, and each item then passed on the retry. Flash never hit the cap.

Three cut-offs in 720 items sounds small, but each one costs more than 150 times what Flash spent on an average everyday answer, and produces nothing. If you run a reasoning model in production, log how often a reply ends at the token limit. Those calls don't show up as errors. They show up as a bigger bill and a retry.

Results by task

Cost per correct answer includes retries. First-try rates are out of 50 items per task, except text_to_sql_hard (30) and regex_from_examples_hard (40).

Intent classification had the smallest gap, 2.3x. The answer is one short label, so there is little room for extra output. The long-document task sends a 30K-token handbook on every call, yet it cost about the same as the short tasks: about 98% of that input was served from DeepSeek's cache at $0.006 or $0.044 per million tokens. If your prompts share a long prefix, caching matters more than which of these two models you pick.

What this does not tell you

• It is one run per model. Two runs would give slightly different token counts. A partial repeat of the hardest invoice task gave V4 Pro 50 out of 50 again, but we have not repeated the whole set.
• Both models scored 100%, so these tasks do not separate them on accuracy yet. Harder tasks might. If V4 Pro solves things Flash can't, the math changes, and the success-rate ratio in the formula above stops being 1.
• Every task has a checkable answer. Open-ended writing, long agent loops and research work are not covered.
• Costs are at Peak list price. Our actual charge was about 46% of that, because most calls ran off-peak at half price. The ratio between the two models is the same either way.
• We used provider defaults. A lower reasoning setting on V4 Pro, if your setup allows one, would cut its token count, and that is worth testing before you rule it out.

The full write-up, per-task token counts and the reworded items are in the results file. Task details and grading rules are in TASKS.md.

Run it on your own models

The runner works with OpenAI, Anthropic, Google Gemini and DeepSeek keys. It prints what a run would cost before making any call, stops before any attempt that could go past your budget, and resumes where it stopped. A full DeepSeek run like this one costs a few dollars at list price.

git clone https://github.com/aicostbudget/llm-cost-per-task-benchmark
cd llm-cost-per-task-benchmark
pip install -r requirements.txt -r requirements-runner.txt

python runner.py --estimate --tier all
export DEEPSEEK_API_KEY=...
python runner.py --run --tier all --models deepseek-flash --budget 1 --run-id flash

The report gives first-try success, success within three attempts, mean attempts and token counts per task. Paste those into the Model Switch Cost Calculator with the success rate source set to measured, and it prices the same comparison at your own volume. For why the cheaper token rate can still lose when success rates differ, read the cheapest LLM can cost more per successful task. The LLM Retry & Fallback Cost Calculator prices a Flash-first route that only sends failures to V4 Pro.

Our own takeaway is narrow. For work where a script can tell you whether the answer is right, start with V4.1 Flash and move a task to V4 Pro only when Flash fails it often enough to show up in your logs.

FAQ