DeepSeek V4-Flash Is the Cheapest Major AI Model on the Market. DeepSeek Itself Is Already Raising Prices

Three cents. That's what it costs to run DeepSeek V4-Flash through the full Artificial Analysis benchmark suite, an independent research firm that ranks AI models by real-world cost, not just raw scores. For comparison, Kimi K3 from Moonshot AI, the model that wiped $3.3 trillion off chipmaker valuations back in July, costs 86 cents on the same test. OpenAI's GPT-5.6 Sol runs $1.86. Anthropic's Claude Fable 5 hits $3.15. DeepSeek V4-Flash isn't just cheaper than its closest big rival, it's cheaper by an order of magnitude, not a few dozen percent.

Cheap, but not the weakest in the room

A low price wouldn't mean much on its own if the model were useless. The July 31 build of V4-Flash scored 50 on Artificial Analysis's Intelligence Index, a composite drawn from nine separate benchmarks covering coding, reasoning, and practical office-style tasks. That's a 10-point jump over the earlier April version of V4-Flash and, more surprisingly, 6 points ahead of the same company's pricier, supposedly more capable DeepSeek V4-Pro, despite sharing identical architecture and pricing. On the coding benchmark specifically, V4-Flash lands at #42 out of 173 ranked models, top quartile, not the top spot, but a result that genuinely competes with models priced many times higher.

The API pricing matches that ambition: $0.14 per million input tokens and $0.28 per million output tokens, with cache-hit input dropping to $0.0028 per million. For a developer building a product on top of high query volume, that's the difference between a bill you have to budget around monthly and one you can more or less ignore.

The twist: DeepSeek is raising its own prices

On August 6, two days before Artificial Analysis crowned V4-Flash the cheapest well-known model on the market, DeepSeek notified developers of an upcoming, significant price increase across its entire API, covering both V4-Flash and V4-Pro. The company points to rising compute costs, capacity bottlenecks, and the sheer traffic volume the model generated after launch. The timing lines up with recent server outages at DeepSeek and plans for a 1-gigawatt data center in Inner Mongolia, so the full picture is consistent: demand outgrew the infrastructure, and raising the price is the cheapest lever available to slow it down.

The second price shock in a month

Kimi K3 from Moonshot AI showed in July that a Chinese model good enough and half the price of Western competitors can wipe billions off chipmaker valuations in two days. DeepSeek V4-Flash goes a step further: it isn't just cheaper, it's cheaper by an order of magnitude than the very model that triggered that earlier shock. For anyone watching the AI market from the sidelines, it's another data point that Chinese model providers are now competing on price as hard as on raw performance, and the gap between top-tier models has narrowed enough that picking one increasingly comes down to cost rather than brand. The same dynamic was behind Alibaba's Qwen3.8-Max launch, which shipped with a full API price sheet and direct comparisons to Western competitors on day one.

The scale of the infrastructure investment DeepSeek announced alongside the price hike explains why the company is reaching for such a drastic move in the first place. A gigawatt of capacity in a single data center is on the order of what a mid-sized city draws, not another rack bolted onto existing infrastructure. Building at that scale costs real billions, and DeepSeek has to fund it from API revenue it deliberately kept artificially low for the past year to grab market share. The recent server outages show the current infrastructure is already struggling to keep up with traffic at pre-hike prices, so this isn't purely a financial problem, it's a technical one too.

The context here says the "cheapest" label has a short shelf life these days. The same DeepSeek that kicked off the entire AI price war with its V3 model in 2025 is now the first major player to officially walk back the lowest rate in the field, before anyone even got used to it. That's not a sign the model stopped being viable, it's a sign the cost of running infrastructure under millions of queries finally caught up to the pricing that got it there in the first place.

What it means for anyone already building on V4-Flash

For teams that already shipped a production integration built on today's V4-Flash pricing, DeepSeek's notice is a cue to re-run the budget before the new rates land, not a reason to panic. The model will likely still be among the cheaper options on the market even after the hike, unless DeepSeek goes for a genuinely drastic jump, which nothing currently suggests. The real risk sits elsewhere: if even the price war's own leader is admitting today's rates don't cover infrastructure costs, the entire category of "models an order of magnitude cheaper than the competition" might turn out to be a transitional phase of the market rather than a stable baseline anyone can safely plan a long-term product budget around.