DeepSeek V4 API Pricing: Peak vs Off-Peak Rates
DeepSeek built its reputation on very low prices, then raised them with V4. The current structure also has different rates at peak and off-peak times.
V4 prices per million tokens
- V4 Flash: $0.44 input / $1.32 output at peak, $0.22 / $0.66 off-peak
- V4 Pro: $1.32 input / $3.96 output at peak, $0.66 / $1.98 off-peak
These are cache-miss input prices. Cache hits cost much less.
How it compares
Even at peak, V4 Flash is far cheaper than mid-tier models like Claude Sonnet 5.5 at $2 / $10. V4 Pro at peak is in the same range as Mistral Large 4 and Grok 4.5.
What to watch
- Price changes: DeepSeek has raised prices sharply, so avoid hard-coding old numbers in your budget.
- Peak pricing: if your workload can run off-peak, you pay half.
- Data location: check where your data is processed and whether that fits your compliance needs.
What V4 costs on a real workload
Take our standard workload of 45 million input and 12 million output tokens a month, and apply the list prices without caching:
| Option | Cost (approx.) |
|---|---|
| V4 Flash at peak | $36 |
| V4 Flash off-peak | $18 |
| V4 Pro at peak | $107 |
| V4 Pro off-peak | $53 |
For comparison, the same workload on Claude Sonnet 5.5 or GPT-6 Sol costs about $210. V4 Pro at peak rates is around half of that, and V4 Flash is a fraction of it. Whether quality is high enough is a separate question that only your own tests can answer.
Peak vs off-peak: can you shift your load?
If part of your workload is not time-sensitive, such as nightly summaries, dataset labeling or bulk translation, scheduling it off-peak halves the price. Check the provider's documentation for the exact peak windows and time zone, then run batch-style jobs outside them. Interactive features that must respond during the day will pay the peak rate.
Which V4 model should you pick?
Flash is the cost-efficient choice for high-volume, simpler tasks, such as classification, extraction and drafting. Pro is the stronger model for harder reasoning and coding, at three times the price. A practical approach is to start on Flash, measure failures on your test set, and move only the failing task types to Pro.
Price history and risk
DeepSeek became known for rock-bottom pricing, then raised its prices sharply with V4. Low-cost providers can change terms quickly, so avoid locking a business plan to today's rate. Keep your integration portable, ideally behind a simple abstraction, so you can switch models if prices or availability change.
Compliance and data questions
Before sending customer or company data, check where processing happens, how long data is kept and whether that meets your legal and contractual requirements. Some organizations prefer to use open-weight models through a third-party host in their own region. DeepSeek has historically published open weights, so hosts may offer it at their own prices, which can differ from the official ones.
Frequently asked questions
Is DeepSeek V4 cheaper than GPT-6 or Claude?
Per token, yes at list prices, especially V4 Flash. Compare cost per successful task on your own data.
Does DeepSeek offer discounts for cached input?
Cache hits are billed at a lower rate than cache misses, so repeated prompts cost less than the list prices above. See prompt caching explained.
Where can I compare it with other models?
Use the cost calculator, which includes both V4 models with editable prices.
Run the numbers
Both V4 models are in the LLM cost calculator at peak rates. Halve the result to estimate off-peak cost, and compare with cheaper small models.
Prices checked November 4, 2026 and subject to change.
Track spend by team and project: the AI Spend Tracker spreadsheet is launching soon.