DeepSeek V4 Pro API Pricing: Peak and Off-Peak Rates

DeepSeek V4 Pro is an open-weight model with a 1M token context window and a hosted API. This page covers price, reported quality, limits and release status. Anything that can change was checked on official pages on 2026-10-15 and carries that date.

The price has two values. At peak hours V4 Pro costs $1.32 per million input tokens and $3.96 per million output tokens. At every other hour it costs half as much.

What the DeepSeek V4 Pro API costs (checked 2026-10-15)

Price per 1M tokens	Peak	Off-peak
Input, cache miss	$1.32	$0.66
Input, cache hit	$0.044	$0.022
Output	$3.96	$1.98

These are DeepSeek's list prices for deepseek-v4-pro, which has a 1M token context and up to 384K output tokens. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays. All other hours are off-peak, including weekends and those holidays. DeepSeek can change prices, so recheck before you commit.

In China time (UTC+8), peak is 09:00 to 12:00 and 14:00 to 18:00. A 09:00 to 17:00 workday in US Eastern time is 13:00 to 21:00 UTC in summer (14:00 to 22:00 in winter), so all of it is off-peak. In central Europe the first two or three hours of the workday are peak. Work that can wait can run off-peak at half price.

For comparison, current GPT prices are in our Azure OpenAI vs OpenAI direct table, Claude's in the Claude API pricing guide and Gemini's in the Gemini 2.5 Pro vs Flash cost guide.

A worked month at each price

Take 500 million tokens in a month: 250 million input tokens (all cache misses) and 250 million output tokens. The even split is an assumption, so rerun the math with yours. Bill = input millions x input price + output millions x output price, using list prices per million tokens as of 2026-10-15, with no caching, batch or volume discounts.

DeepSeek V4 Pro

• Peak: 250 x $1.32 + 250 x $3.96 = $330 + $990 = $1,320.
• Off-peak: 250 x $0.66 + 250 x $1.98 = $165 + $495 = $660.
• Even traffic all week: peak is 7 hours a day on 5 days, which is 35 of the 168 weekly hours. (35 x $1,320 + 133 x $660) / 168 = ($46,200 + $87,780) / 168 = $797.50. Chinese public holidays would lower this a little.

Other models, same tokens

• Claude Opus 5.5 ($4 input, $20 output): 250 x $4 + 250 x $20 = $1,000 + $5,000 = $6,000.
• Claude Opus 4.7 ($5, $25): 250 x $5 + 250 x $25 = $1,250 + $6,250 = $7,500.
• GPT-5.5 ($5, $30, prompts up to 272K tokens): 250 x $5 + 250 x $30 = $1,250 + $7,500 = $8,750.

What moving traffic saves

Moving every token from Opus 4.7 to V4 Pro at peak saves $7,500 - $1,320 = $6,180 a month, or $6,180 x 12 = $74,160 a year. Against Opus 5.5 it saves $6,000 - $1,320 = $4,680 a month. If only a share of your tokens moves, multiply by that share: 60% moving from Opus 4.7 saves 0.6 x $6,180 = $3,708 a month, or $44,496 a year. The 60% is an input, not a forecast. At 5 billion tokens a month, every figure is ten times larger.

What changes these numbers

• Tokenizers differ. Each vendor counts tokens its own way. Anthropic says the tokenizer in Claude 4.7 and later produces about 30% more tokens for the same text than the one in Claude Sonnet 4.6 and earlier.
• Thinking mode. DeepSeek's documentation says thinking is on by default at high effort. Check the output token counts in your own responses before you trust a 50/50 split.

What DeepSeek reported about quality

Every number in this section is DeepSeek-reported: DeepSeek ran the tests and published them on its own model cards, so none of them is an independent measurement.

On the April 24 preview card, DeepSeek reports V4-Pro (Max setting) at 80.6 on SWE Verified, 67.9 on Terminal Bench 2.0, 93.5 on LiveCodeBench and 3,206 on the Codeforces rating. The same card lists three comparison models:

• Claude Opus 4.6 (max): 80.8, 65.4 and 88.8 on the first three tests.
• GPT-5.4 (xHigh): 75.1 on Terminal Bench 2.0 and 3,168 on Codeforces.
• Gemini 3.1 Pro (high): 80.6, 68.5 and 91.7 on the first three tests, and 3,052 on Codeforces.

On that table V4-Pro sat 0.2 below Opus 4.6 on SWE Verified and trailed GPT-5.4 on Terminal Bench 2.0 by 7.2 points (75.1 - 67.9). It led every listed model on LiveCodeBench, where Gemini 3.1 Pro was next at 91.7 (a 1.8 point gap), and on Codeforces, where GPT-5.4 was next (a 38 point gap, 3,206 - 3,168). The card does not include Opus 4.7 or any later Claude model, so it cannot tell you how V4 Pro compares with them.

Those scores are for the preview build. The general availability build, V4-Pro-0813, shipped on 2026-08-13 and replaces the preview. Its card moved to Terminal Bench 2.1 and lists 87.9 for V4-Pro-0813 against 72.1 for the preview, with Claude Opus 4.8 at 85.0. DeepSeek ran those agent tests with its own harness, and 2.1 is a different test from 2.0, so do not set 87.9 beside 67.9.

A vendor-reported score is a reason to test, not a result. Run your own prompts before you move traffic.

Release status (checked 2026-10-15)

• 2026-04-24: V4 launched as a preview, with weights on Hugging Face under the MIT license. V4-Pro has 1.6T total and 49B active parameters.
• 2026-08-13: V4-Pro reached general availability under the same name, deepseek-v4-pro, with three thinking effort levels (low, high, max). Peak and off-peak pricing started on 2026-08-16.
• 2026-09-10: DeepSeek released V4.1 Flash as deepseek-flash (listed at $0.30 cache-miss input and $1.20 output at peak, half that off-peak) and retired the earlier V4 Flash. It also said V4 Pro was being phased out, and that from September 14 requests to deepseek-v4-pro would go to V4.1 Flash at Flash prices until a V4.1 Pro launches.
• Now: DeepSeek's change log says it will keep serving V4 Pro after September 14 with billing unchanged, and will give further notice of any change.

V4 Pro is available today at the prices above, but DeepSeek has signaled a successor. Pin the model name, read the change log before a large rollout, and keep a second model you have tested. The provider failover guide covers the setup.

Limits to plan for (checked 2026-10-15)

DeepSeek allows 500 concurrent requests per account for deepseek-v4-pro and 2,500 for deepseek-flash, and returns HTTP 429 beyond that. A capacity expansion request costs nothing, according to DeepSeek. The server closes a request that has not started inference after 10 minutes.

Self-hosting the open weights

The weights are public under the MIT license, and they are large: V4-Pro has 1.6T total parameters. vLLM's recipe for it lists checkpoints of several hundred gigabytes, and on 8 H200 GPUs it caps context at 800K tokens, below the API's 1M (checked 2026-10-15). Whether self-hosting beats the API is a break-even calculation. Gemma 4 31B: Self-Host vs API Routing Math shows the method with GPU prices and utilization, and the table above is the number to beat. Our GLM-5.1 post covers a different coding model.

Where PromptUnit fits

PromptUnit is a proxy: you point your app's AI API base URL at it, and it checks each request. A cheaper model is used only when it is just as good for that request. Otherwise the request goes to the model you asked for. It works with OpenAI, Anthropic, Google, Groq and DeepSeek keys, and with your own OpenRouter key. The dashboard shows each request: the model you asked for, the model that answered, the cost and the saving.

From our early data on one production account: about 83% of live requests were simple question-answering and about 6% were heavy reasoning or complex code. In 24 side-by-side comparisons a cheaper model matched or beat the chosen model 18 times (75%), at roughly 89% lower cost on those calls. It is a small sample, so check your own mix first (method: What We Measured Routing GPT-4o Traffic to Cheaper Models).

To compare models yourself, send the same prompts to each one with the header x-promptunit-routing: off, which keeps a request on the model you asked for, and read the costs side by side in the dashboard. For the general method, see LLM Model Routing: The Complete Guide. For coding picks by task, see Best LLM for Coding in 2026.

PromptUnit is free until it has saved you $5 (individual) or $100 (business, work email). After that it takes 20% of savings. Start free or talk to us.

Changes in this version

Updated 2026-10-07: added off-peak and cache-hit prices, fixed the monthly arithmetic and the benchmark comparisons, and added release status and limits.