DeepSeek V4 API Pricing: Official vs OpenRouter vs DeepInfra vs cloudgpu.app

“What does DeepSeek V4 cost per million tokens” has at least five correct answers on any given day, because the same model id is sold by DeepSeek itself, by aggregators such as OpenRouter, by hosts running the open weights such as DeepInfra and Together, and by gateways such as ours. The numbers differ by a factor of ten and the model behind the name is not always the same. This post puts them side by side as read on 10 September 2026, explains what a gateway adds and what it costs, and is blunt about the rows where cloudgpu.app is not the cheapest option, because today there are several.

The table

Per million tokens, US dollars, as displayed on each provider’s public pricing page or API on 10 September 2026. “Cached” is the price for a prompt-cache hit where the provider publishes one.

deepseek-v4-flash

Provider	Checkpoint served under this name	Input	Cached input	Output
DeepSeek official (deepseek-flash; legacy id deepseek-v4-flash accepted)	V4.1 Flash	$0.30 peak / $0.15 off-peak	$0.006 / $0.003	$1.20 / $0.60
OpenRouter deepseek/deepseek-v4-flash	V4 Flash 0423	$0.0886	$0.0177	$0.1772
OpenRouter deepseek/deepseek-v4.1-flash	V4.1 Flash	$0.15	$0.003	$0.60
DeepInfra DeepSeek-V4-Flash	V4 Flash	$0.09	$0.018	$0.18
Together	V4 Flash 0731 only	$0.14	$0.03	$0.28
cloudgpu.app deepseek-v4-flash	routed to DeepSeek or a partner host, see below	$0.396	$0.198	$1.188

deepseek-v4-pro

Provider	Checkpoint served under this name	Input	Cached input	Output
DeepSeek official deepseek-v4-pro	V4 Pro 0813 (routed to V4.1 Flash from 14 Sep 2026)	$1.32 peak / $0.66 off-peak	$0.044 / $0.022	$3.96 / $1.98
OpenRouter deepseek/deepseek-v4-pro	V4 Pro 0423	$0.9553	$0.0796	$1.9105
OpenRouter deepseek/deepseek-v4-pro-0813	V4 Pro 0813	$0.6587	$0.021	$1.976
DeepInfra DeepSeek-V4-Pro	V4 Pro	$1.30	$0.10	$2.60
Together	V4 Pro 0813	$1.32	$0.13	$3.96
cloudgpu.app deepseek-v4-pro	routed to DeepSeek or a partner host	$1.3065	$1.3065	$3.564

Sources: DeepSeek’s Models & Pricing page; OpenRouter’s public /api/v1/models feed and model pages; DeepInfra pricing; Together pricing; our own pricing feed. OpenRouter lists prices per token to many decimals; we rounded to four. DeepSeek’s peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; every other hour is off-peak at half price.

Reading the table honestly

Where we stand on Flash. We advertise “at or under the vendor list price”, and for Flash that is not true today. DeepSeek repriced when it replaced V4 Flash with V4.1 Flash: $0.30 input and $1.20 output at peak, half that off-peak, and cache hits at $0.006. Our $0.396 / $1.188 was set against the previous list of $0.44 / $1.32, which is still the “vendor list” figure in our pricing feed. We are one cent under the official peak output price and above it on input, on cached input by a wide margin, and at every off-peak hour. The third-party hosts are cheaper still because they serve the older 0423 checkpoint from open weights. Our Flash price is being reviewed against the new list; until it changes, if Flash is your main model and you can pay DeepSeek directly, DeepSeek is cheaper.

Where we stand on Pro. Against the official peak price we are under on both input ($1.3065 vs $1.32) and output ($3.564 vs $3.96), and under Together on output. OpenRouter’s providers and DeepInfra are cheaper than we are, and DeepSeek’s own off-peak rate is half ours. Our cached-input price for Pro is the same as uncached, which is a real gap against every other row.

The name is not the model. On OpenRouter and DeepInfra, deepseek-v4-flash is the April checkpoint; on DeepSeek’s platform it now means V4.1 Flash. On our gateway the request goes to DeepSeek’s official API or to a partner host running the same open weights, with failover between them, so which checkpoint answers depends on the route. If a specific checkpoint matters to you, use a provider that names it (deepseek-v4-flash-0731, deepseek-v4-pro-0813) and pin it.

Pro is going away at the source. DeepSeek’s page states that from 12:00 Beijing time on 14 September 2026, deepseek-v4-pro requests on its platform are routed to V4.1 Flash and billed at the Flash price, until a V4.1 Pro ships. Hosts running the open Pro weights are not affected. If your product depends on Pro-class reasoning, decide now which of those you want.

What a gateway adds

None of the above says a gateway is pointless; it says you should know what you are paying for.

• One balance for GPUs and APIs. The same prepaid balance that calls DeepSeek here also rents an RTX 4090D at $0.32/h or an RTX 5090 at $0.59/h (live inventory, 10 September 2026). Prototype against the API, move the heavy path to vLLM on a rented card, one ledger.
• A payment method that works without a US card. USDT on TRC-20, minimum 1 USDT, no fee on our side; bank transfer for business customers; card payments are coming soon. Sign-up is an email or a Google account. DeepSeek’s platform has no top-up fee but its checkout rejects many non-Chinese cards. OpenRouter accepts cards and crypto and charges 5.5 percent (minimum $0.80) on card top-ups and 5 percent on crypto, then passes provider prices through unmarked.
• An OpenAI-compatible endpoint with failover. https://cloudgpu.app/v1 speaks the OpenAI chat completions format, so Cursor, Cline, Continue, LangChain and the official SDKs need only a base_url. When the primary upstream times out, the call is retried on the second supplier.
• Chinese vendors on one key. GLM-5.3, Kimi K3, MiniMax M3 and FLUX sit behind the same key at their own prices, listed on the API page.

What it does not add: a lower token price than the cheapest open-weights host, or data residency. Our gateway is in Hong Kong; DeepSeek’s API and one partner host are in mainland China, another partner is in the US, and failover can move a request between them. Do not send data your contracts say may not leave your jurisdiction.

The Python snippet

from openai import OpenAI

client = OpenAI(base_url="https://cloudgpu.app/v1", api_key="YOUR_CLOUDGPU_KEY")

resp = client.chat.completions.create(
    model="deepseek-v4-flash",          # or "deepseek-v4-pro"
    messages=[
        {"role": "system", "content": "You are a terse assistant."},
        {"role": "user", "content": "Explain prompt caching in three sentences."},
    ],
    stream=True,
)
for chunk in resp:
    print(chunk.choices[0].delta.content or "", end="", flush=True)
print()

The response includes a usage object with prompt and completion token counts, and cached prompt tokens where the upstream reports them, so you can reconcile a bill against the table above.

A worked month

Take an agent that sends 40 million input tokens and 8 million output tokens a month on Flash, with 60 percent of input hitting the prompt cache, all during peak hours to keep it simple.

Provider	Uncached input (16M)	Cached input (24M)	Output (8M)	Month	Top-up fee
DeepSeek official (V4.1 Flash)	$4.80	$0.14	$9.60	$14.54	none, if your card is accepted
OpenRouter deepseek-v4.1-flash	$2.40	$0.07	$4.80	$7.27	+5% crypto ($0.36) or 5.5% card ($0.80 minimum)
DeepInfra DeepSeek-V4-Flash (0423)	$1.44	$0.43	$1.44	$3.31	card
cloudgpu.app deepseek-v4-flash	$6.34	$4.75	$9.50	$20.59	none on USDT

At that volume the gateway premium over going direct is about six dollars a month, and DeepInfra’s older checkpoint is a fifth of our price. Whether six dollars buys you a working payment method and one bill is your call; we would rather you make it with the numbers in front of you than find out from the invoice.

If a competitor’s price in this post is wrong, email support@cloudgpu.app with the source and we will correct it the same day. Per-token rates are on the API page. Prices here are a snapshot from 10 September 2026; the pricing page is live.