DeepSeek V4 Flash: Pricing, Benchmarks, and Cost

What is DeepSeek V4 Flash?

DeepSeek V4 Flash is DeepSeek’s efficiency tier: a 284-billion-parameter Mixture-of-Experts model with 13 billion active parameters and a 1M-token context window. The current build, DeepSeek-V4-Flash-0731, shipped on July 31, 2026 as an API public beta and scores 50 on the Artificial Analysis Intelligence Index at max reasoning effort.

That last sentence is doing more work than it looks. There are two models wearing this name, and almost every page describing “DeepSeek V4 Flash” right now is describing the older one.

How much does DeepSeek V4 Flash cost?

Straight from DeepSeek’s own price page, per million tokens:

Rate	DeepSeek V4 Flash	DeepSeek V4 Pro
Input (cache miss)	$0.14	$0.435
Input (cache hit)	$0.0028	$0.003625
Output	$0.28	$0.87
Context	1M tokens	1M tokens

The cache-hit rate is the number worth staring at. At $0.0028 per million tokens, a cached prefix costs about one fiftieth of a fresh read. If you’re running an agent that re-sends the same system prompt and the same repository context on every turn, most of your input bill can disappear. That is a structurally different cost shape from the roughly 90 percent cache discounts Anthropic, OpenAI and Google offer, and it’s a large part of why DeepSeek keeps showing up at the bottom of the AI model price tracker.

One caveat that nobody else seems to be flagging. DeepSeek’s price page also states that prices will be 2x during peak hours, defined as 09:00 to 12:00 and 14:00 to 18:00 Beijing time, with the effective date still pending an official announcement. If that lands, a workload scheduled inside those windows costs double the table above. Worth building into any forecast now rather than discovering it later.

Why you will see four different prices for this model

Search for the price and you’ll get $0.14, $0.10, $0.09, and $0.06, all presented as fact. They’re all real. They’re measuring different things.

DeepSeek V4 Flash input and output price per million tokens, by seller

Item	Output	Input
DeepSeek (first party)	$0.28	$0.14
DeepInfra (resale)	$0.20	$0.10
OpenRouter (resale)	$0.18	$0.09

The fourth number, the one near $0.06 or $0.18, is not a per-token rate at all. It’s Artificial Analysis’s blended cost figure, which folds an assumed input-to-output ratio into a single dollar value. Comparing it to a rate card is a category error, and it’s the single most common mistake in the aggregator pages ranking for this term.

Which checkpoint are you actually calling?

Here’s the part that matters if you’re reading benchmark numbers.

DeepSeek shipped V4 Flash on April 24, 2026. On July 31 it shipped DeepSeek-V4-Flash-0731, described in DeepSeek’s API changelog as a re-post-training of the same 284B architecture. Same model id. Same price. Same parameter count. Very different scores.

So when a spec page tells you DeepSeek V4 Flash scores 40, or 37, or 29, it isn’t necessarily wrong. It’s quoting the April build, and possibly at a different reasoning effort. On that build, Artificial Analysis measured three configurations:

Reasoning effort	April build	0731 build
Non-think	29	not yet published
Think High	37	not yet published
Think Max	40	50

An unlabelled single score for this model is close to meaningless. Ask which build and which effort mode before you trust one.

DeepSeek V4 Flash vs V4 Pro

The interesting result of the 0731 update isn’t that Flash improved. It’s that Flash passed the model it was supposed to sit beneath.

Six points of Intelligence Index, at roughly a third of the token price. And the eval cost corroborates it: Artificial Analysis spent $72.02 running its Index on Flash 0731 against $176.34 on V4 Pro. That figure is useful precisely because it isn’t a sticker rate. It’s what one fixed basket of real work actually cost, verbosity and all.

Which leaves V4 Pro in an odd spot. It’s larger, slower, three times the price, and now scores lower. Its remaining edge is that its weights are published. If you’re picking between the two on the API today, the choice is not close.

How it compares to GPT-5.6 Luna and GLM-5.2

At 50, V4 Flash 0731 sits one point below both GPT-5.6 Luna at max effort and Zhipu’s GLM-5.2, each on 51. For practical purposes that’s a tie: the Intelligence Index blends nine separate evaluations, and a single point is inside the noise you’d expect between runs.

The prices are not a tie. Weighting input and output 3 to 1, the mix that agentic coding actually produces, the blended rates come out like this:

Model	Intelligence Index	Blended $/Mtok	Index points per dollar
DeepSeek V4 Flash 0731	50	$0.175	286
GPT-5.6 Luna	51	$0.45	113
DeepSeek V4 Pro	44	$0.544	81
GLM-5.2	51	$2.15	24

So Flash buys roughly 2.5 times the measured intelligence per dollar that Luna does, and about twelve times what GLM-5.2 does. That makes it the top intelligence-value row on the AI model value leaderboard, by a margin nothing else on the board comes close to.

Worth remembering what this metric does and doesn’t say. Points per dollar rewards cheapness. It doesn’t tell you whether a model can finish your hardest task at all, and there’s no coding composite published for Flash 0731 yet, so the coding picture is genuinely unknown rather than merely unflattering. For a head-to-head that does have coding numbers on both sides, see Kimi K3 versus DeepSeek V4.

Is DeepSeek V4 Flash open source?

Partly, and the distinction matters more than usual this week.

The V4 Flash family is released under the MIT license, and the Hugging Face model card confirms the 284B total and 13B active parameter counts. That’s about as permissive as licensing gets, and it’s why DeepSeek keeps appearing in the best open-weight AI models roundup.

But the card carried no 0731 revision at the time of writing, and DeepSeek describes the 0731 release as an API public beta. So the weights you can download are the April build that scores 40, and the checkpoint that scores 50 is available through the API only. Anyone calling that 50 an open-weights record is, for now, wrong. If DeepSeek posts the 0731 weights under MIT, it becomes the strongest open-weight score published. That hasn’t happened yet.

Should you use it?

Three rules, in the order they actually bite.

Use it if your workload is input-heavy and cache-friendly. Long system prompts, large repository context, many turns over the same material. The $0.0028 cache-hit rate is where the real saving lives, not the $0.14 sticker.

Don’t use it as a drop-in for V4 Pro on knowledge-heavy work without testing. Thirteen billion active parameters is a real constraint no amount of thinking budget fully removes. The Index says Flash is ahead overall; that’s an average across nine evaluations, not a promise about your specific task.

Don’t build a compliance story on the weights. If your reason for choosing DeepSeek is self-hosting or open weights, you’re getting the April build today. Check Hugging Face before committing.

And if a Chinese-lab model is off the table for procurement reasons rather than technical ones, that’s a separate question with a separate answer, covered in the China AI pricing breakdown.

Frequently asked questions

How much does DeepSeek V4 Flash cost?

DeepSeek charges $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens. Resellers price it lower: OpenRouter lists $0.09 and $0.18, DeepInfra $0.10 and $0.20. A pending 2x peak-hour surcharge is stated on DeepSeek price page but has no announced start date.

Is DeepSeek V4 Flash better than V4 Pro?

On the Artificial Analysis Intelligence Index, yes. The 0731 build of V4 Flash scores 50 at max reasoning effort against 44 for V4 Pro, at roughly a third of the token price, and Artificial Analysis spent $72.02 evaluating Flash against $176.34 for Pro. V4 Pro keeps two advantages: far more active parameters for knowledge-heavy work, and published weights.

What changed in the DeepSeek V4 Flash 0731 update?

DeepSeek re-post-trained the same 284B architecture and shipped it on July 31, 2026 behind the same model id and the same price. The Intelligence Index score rose 10 points to 50 at max effort. DeepSeek reports large agent-benchmark gains, including Terminal Bench 2.1 at 82.7, and added native Responses API support.

Is DeepSeek V4 Flash open source?

The V4 Flash family is MIT licensed and the April 2026 build is downloadable from Hugging Face. The 0731 checkpoint that scores 50 had not been posted at the time of writing and is available through the API only, so the downloadable weights and the benchmark-leading model are not currently the same thing.

What is the DeepSeek V4 Flash context window?

One million tokens, with a 384K maximum output. That matches V4 Pro. The model uses a hybrid attention design intended to keep long-context inference cheap, which is the main reason its per-token price can sit this far below comparable models.

Sources

• DeepSeek. (2026). Models and pricing [first-party per-token rates, cache-hit rates, pending peak-hour surcharge]. Verified 2026-07-31. api-docs.deepseek.com
• DeepSeek. (2026). Change log [0731 release date, public-beta status, vendor-reported agent benchmarks, Responses API support]. Verified 2026-07-31. api-docs.deepseek.com
• Artificial Analysis. (2026). DeepSeek V4 Flash 0731: Intelligence, Performance and Price Analysis [Intelligence Index 50 at max effort, $72.02 cost to run the Index, release date]. Verified 2026-07-31. artificialanalysis.ai
• Artificial Analysis. (2026). DeepSeek V4 Pro: Intelligence, Performance and Price Analysis [Intelligence Index 44 at max effort, $176.34 cost to run the Index, 1.6T total and 49B active parameters, MIT license]. Verified 2026-07-31. artificialanalysis.ai
• DeepSeek. (2026). DeepSeek-V4-Flash model card [MIT license, 284B total and 13B active parameters, 1M context]. Verified 2026-07-31. huggingface.co
• OpenRouter. (2026). DeepSeek V4 Flash [reseller rate $0.09/$0.18]. Verified 2026-07-31. openrouter.ai