DeepSeek V4 Flash Review: Cost per Benchmark Point vs V4 Pro
Pick the capability you care about, enter your monthly token volume, and this DeepSeek V4 Flash review divides each model's monthly bill by its published score on that benchmark, then tells you what V4 Pro would have to score on that same benchmark to reach the same cost per point.
Updated 2026-08-05, by Jim Liu
TL;DR
- DeepSeek-V4-Flash-0731 shipped 2026-07-31, weights on Hugging Face under MIT, API in public beta the same day
- List price is $0.14/M input on a cache miss, $0.0028/M on a cache hit, $0.28/M output
- V4 Pro lists at 3.11x Flash on the cache-miss input meter and 3.11x on the output meter. Its cached-input rate is not at that ratio, so the multiple below holds only for cache-miss pricing
- Flash-0731 scores above V4-Pro (Preview) on all nine benchmarks DeepSeek published with the release
- Those nine scores are DeepSeek's own, measured against a preview build of Pro. We did not rerun them
Cost per benchmark point: V4 Flash against V4 Pro
Prompts, tool schemas and retrieved files, priced at the cache-miss rate.
Completion tokens, as billed by the API.
On Terminal Bench 2.1
3.11x the price of Flash
The V4-Pro list rate is 3.11 times Flash's: $109 a month against $35.00. The Pro Preview build also scores 10.6 points lower on this benchmark, so Pro would need to score 257.0 here to match Flash on cost per point, which is off a 100-point scale.
$0.4232 per benchmark point per month · 0731 release score
$1.51 per benchmark point per month · Preview build score, current list price
Pro costs $73.75/mo more at 200M input and 25M output tokens. Cache-miss input pricing, list rates as of 2026-08-05.
Try a preset workload
All Nine Published Benchmarks, Side by Side
| Benchmark | V4-Flash-0731 | V4-Pro (Preview) | V4-Flash (Preview) | Gain vs Flash preview |
|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 72.1 | 61.8 | +20.9 |
| NL2Repo | 54.2 | 38.5 | 39.4 | +14.8 |
| Cybergym | 76.7 | 52.7 | 38.7 | +38.0 |
| DeepSWE | 54.4 | 12.8 | 7.3 | +47.1 |
| Toolathlon-Verified | 70.3 | 55.9 | 49.7 | +20.6 |
| Agents' Last Exam | 25.2 | 16.5 | 15.8 | +9.4 |
| AutomationBench Public | 25.1 | 12.8 | 10.8 | +14.3 |
| DSBench-FullStack | 68.7 | 41.8 | 37.0 | +31.7 |
| DSBench-Hard | 59.6 | 31.1 | 25.8 | +33.8 |
The last column is where the release story sits. DeepSeek describes the architecture and parameter count as unchanged between the Flash preview and the 0731 build, attributing the movement to a new post-training pipeline. DeepSWE moves from 7.3 to 54.4 at a fixed model size, the largest such movement on the table, and a reason not to assume the jump generalises to workloads nobody benchmarked.
What the Checker Divides
A score and a price sit on unrelated scales: 82.7 against 72.1 on one side, $0.28 against $0.87 per million tokens on the other. The checker collapses them into one unit, dollars per benchmark point per month, by taking a single benchmark and dividing the monthly bill by the published score on it.
The bill is a multiplication. Input tokens are multiplied by the cache-miss input rate, output tokens by the output rate, and the two are added. Every input token is priced as a cache miss on purpose. DeepSeek does discount a repeated prompt prefix heavily, down to $0.0028 per million on V4 Flash, but how much of your prompt repeats depends on how you build requests, and DeepSeek does not publish which prefixes qualify for a hit or how long one is kept. Cache-miss pricing is the one assumption that needs no guess about your hit rate, and the two models are much closer together on the cached-input meter. To model that separately, the prompt caching cost calculator covers cache-hit economics and the break-even hit rate in detail.
Each of the nine axes measures something different enough that averaging them would destroy the information. Cybergym is a security suite. DSBench-FullStack builds working data applications. Toolathlon measures whether tool calls come back correctly across many turns. A model can score well on one of those and poorly on another, which is why the checker makes you pick an axis instead of handing you a composite figure. Dividing by a score also treats benchmark points as if they were linear, which they are not: a model scoring 60 is not twice as useful as one scoring 30. Cost per point is a way to compare two models on one axis, not a measure of capability per dollar in any absolute sense.
The break-even score answers a question the price table leaves open: given how much more V4 Pro costs, how far ahead would it have to score before the two models cost the same per point? At a 3.11x cost multiple, Pro would need roughly triple Flash's score to reach the same cost per point. On Terminal Bench 2.1 that means about 257, on an axis whose two published scores are 82.7 and 72.1 and which is reported out of 100. At list prices with no cache hits, Pro cannot reach break-even there. A cut to Pro's price or a high cache hit rate would both lower that target.
What the checker will not do is rank Flash against models outside the DeepSeek family. Cross-vendor benchmark numbers are not measured under the same harness, so a ranking built from them would not mean anything. For a straight per-token price comparison across vendors, use the LLM API token cost calculator instead.
Why the Cost Multiple Barely Moves With Your Token Mix
Comparisons between two models normally depend on your token mix. A model with cheap input and expensive output wins on document reading and loses on long generation, so the honest answer to "which is cheaper" is usually "it depends how much you write back". Not here, at least under the pricing this checker uses.
V4 Pro lists at $0.435/M input against Flash's $0.14/M, a ratio of 3.107. On output it lists at $0.87/M against $0.28/M, a ratio of 3.107. Identical to three decimal places, so the published rates are proportional across both meters. Because that factor is common to every term, it cancels out of the ratio, so the multiple in the checker stays at 3.11x whether your workload is 4,000M input tokens with almost no output, or the reverse.
The exception is the cached-input meter, where V4 Pro sits at $0.003625/M against Flash's $0.0028/M, a ratio of about 1.29. A workload with a very high cache hit rate therefore narrows the gap. If both models see a similar share of cache hits, a high-hit workload lands between the two ratios; if their hit shares differ, it can land outside that range, and nothing published lets us tell which from here.
So under this checker's all-cache-miss assumption, the token mix changes your absolute bill, sometimes by a lot, without changing which model is cheaper or by how many times. Introduce cache hits and the multiple does move, toward the 1.29x figure above. For the history of how Pro reached its current rates, our earlier hands-on write-up of the DeepSeek V4 Pro price cut walks through the bill before and after that change.
Where These Numbers Come From, and Where the Sources Disagree
We did not run any of these benchmarks. Doing it credibly means reproducing nine harnesses against a beta API, which we have not done. Every score here is DeepSeek's own published figure, read from the model card on 2026-08-05, and every price is the list rate from DeepSeek's pricing page read the same day. What we did was redo the arithmetic ourselves and record where the published sources conflict with each other.
One conflict shows up in almost every secondary write-up without comment. OpenRouter's model page describes V4 Flash as 13B active parameters out of 284B total. The Hugging Face model card states 304B total parameters. Both cannot be right as stated, and we do not know which is. Neither figure changes the pricing or the benchmark results, so the page carries both.
- Benchmark scores and license: DeepSeek-V4-Flash-0731 model card
- API list pricing: deepseek.ai/pricing
- Parameter counts, context window, third-party hosting rates: OpenRouter model page
- Throughput and latency by provider: Artificial Analysis provider table
- Data as of: 2026-08-05. Prices and beta status change; verify before budgeting
Jim Liu
Builds and ships AI tooling, and writes token-budgeting and cost-optimization guidance for teams running production LLM pipelines. Publishes tools and analysis at OpenAI Tools Hub.
Questions People Ask Before Switching
Is DeepSeek V4 Flash free?
The weights are free, the API is not. DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face under an MIT license on 2026-07-31, so you may download and self-host it at your own hardware cost. The hosted API is metered: $0.14 per million input tokens on a cache miss and $0.28 per million output tokens as of 2026-08-05.
How does DeepSeek cache pricing work on V4 Flash?
DeepSeek bills a repeated prompt prefix at $0.0028 per million tokens instead of $0.14, so that portion of the input costs one fiftieth of the cache-miss rate. The published rates list no separate cache-write charge, though they do not describe which prefixes qualify for a hit or how long one is retained. The checker on this page prices every input token at the cache-miss rate, because it does not require guessing a hit rate. On the cached-input meter the two models are only about 1.29x apart rather than 3.11x, so if both see a similar share of cache hits, a high-hit workload narrows the gap.
Is DeepSeek V4 Flash better than V4 Pro?
On the nine agent and coding benchmarks DeepSeek published with the 0731 release, V4-Flash-0731 scores higher than V4-Pro (Preview) on all nine, while listing at roughly a third of the price on the cache-miss input and output meters. Those are DeepSeek's own numbers on its own release, and they compare against a preview build of Pro rather than a final one. What they support is the narrow statement that Flash scored higher in these nine published results, not a general ranking.
What changed between the Flash preview and the 0731 release?
DeepSeek describes the parameter count and architecture as unchanged and attributes the gains to a new post-training pipeline. The clearest illustration is DeepSWE, where the reported score moves from 7.3 on the preview to 54.4 on the 0731 build at an unchanged reported model size. The published data does not independently isolate the cause, so this remains DeepSeek's account of its own release.
How big is DeepSeek V4 Flash, and how much of it runs per token?
It is a sparse mixture-of-experts model with a 1M token context window. OpenRouter's model page lists 13B active parameters out of 284B total; the Hugging Face model card states 304B total parameters. We cite both because we cannot tell which is right.
Related Tools
LLM API Token Cost Calculator
Per-token pricing across vendors for your exact prompt and reply length
Prompt Caching Cost Calculator
Model the cache-hit economics this page deliberately leaves out
Context Window Calculator
See what actually fits in a 1M token window before you plan around it
AI Model Cost Comparison
Three questions, then a ranked pick across Claude, GPT, Gemini and DeepSeek