DeepSeek API Pricing (November 2026): V4.1 Flash and V4-Pro Costs, Off-Peak Discount and “Server Is Busy” Fix

DeepSeek is still the price leader among serious AI APIs, and in autumn 2026 it got simpler to use. There are now just two model names: deepseek-flash, which runs the new DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, which runs DeepSeek-V4-Pro-0813. Both have a 1M token context and up to 384K output tokens. DeepSeek also moved to a time-of-day price: everything is half price outside its peak hours.

In this guide I cover the official prices (from DeepSeek’s API docs, checked November 5, 2026), how the peak and off-peak windows work, what the limits are, and how to deal with the infamous “server is busy” message in both the app and the API.

DeepSeek API Prices

Both models have a 1M token context length and a maximum output of 384K tokens, which is three times what Claude and GPT-6 allow on their normal synchronous endpoints.

If you used DeepSeek earlier in 2026, note that the Flash price dropped. When V4-Flash 0731 launched, Flash cost $0.44 input and $1.32 output at peak. The new deepseek-flash name costs $0.30 and $1.20. DeepSeek says the old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but their original models are retired, so those requests are now served by V4.1-Flash-1017 and billed at the new Flash price.

DeepSeek’s models are also open weights under the MIT licence, so you can run them yourself or through other providers. I compare DeepSeek with Claude in detail, including running V4-Flash on a 128 GB RAM PC, in DeepSeek vs Claude.

Peak vs Off-Peak Hours

DeepSeek charges full price only during two weekday windows and half price the rest of the time:

That means DeepSeek is at the off-peak price for most of the week: 7 peak hours a day on weekdays, so about 35 of the 168 hours in a week. If you run batch jobs, scheduling them outside those windows halves your bill with no code changes. Time zone conversions above use standard time; check again when daylight saving changes.

Context Caching: Why Cache Hits Are Almost Free

DeepSeek caches repeated prompt prefixes automatically. When the start of your prompt matches something it has seen recently, those tokens are billed at the cache hit price, which is 2% of the cache miss price on Flash ($0.006 vs $0.30 at peak) and about 3% on V4-Pro. You do not need to mark anything in the request.

To get the most hits, keep the stable parts of the prompt (system prompt, instructions, reference documents, few-shot examples) at the very beginning and identical across requests, and put the changing user message at the end.

Cost Examples

The same support bot workload I use in my other API guides: 90M input and 12M output tokens a month.

For comparison, the same workload costs $300 on Claude Sonnet 5.5 or GPT-6.1 Sol, $252 on Grok 4.7 and $112.50 on Gemini 3.8 Flash. Only Claude Haiku 5.5 and GPT-6 Luna (both about $15) come in cheaper than DeepSeek Flash for short prompts. See my AI API pricing comparison for the full table. All figures are estimates from list prices.

Concurrency Limits and Timeouts

DeepSeek does not use requests-per-minute limits like OpenAI or Anthropic. Instead it limits concurrent requests per account, across all your API keys:

• deepseek-flash: 2,500 concurrent requests.
• deepseek-v4-pro: 500 concurrent requests.

A request occupies a slot from the moment it is sent until the model finishes answering. Go over the limit and you get HTTP 429. You can ask for more through DeepSeek’s capacity expansion form at no extra cost. Passing a user_id does not give you extra capacity on a standard account; all user IDs share the same pool.

While DeepSeek is preparing a response, it keeps the connection open: non-streaming requests receive empty lines and streaming requests receive SSE : keep-alive comments. The SDKs ignore these, but if you parse raw HTTP yourself, skip them. If inference has not started after 10 minutes, the server closes the connection.

DeepSeek API Error Codes

How to Fix “Server Is Busy”

“The server is busy. Please try again later.” is DeepSeek’s most famous message. In the API it corresponds to 503 Server Overloaded, which DeepSeek describes as the server being busy due to high traffic. Here is what actually helps:

In the API

1. Retry with exponential backoff and jitter (wait 1, 2, 4, 8 seconds plus a random fraction). The OpenAI SDK can do this for you with max_retries.
2. Use streaming so you get keep-alive signals and the first tokens as soon as possible.
3. Lower your parallelism. If you see 429 rather than 503, you are over your concurrency limit, not hitting a DeepSeek-wide problem.
4. Run big jobs off-peak. Outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, load is typically lower and the price is half.
5. Add a fallback. Because DeepSeek’s weights are MIT licensed, the same models are offered by other inference providers, and you can also fall back to a different model family.

In the DeepSeek app or website

1. Wait a minute and press regenerate rather than sending the message again.
2. Start a new chat. Very long conversations take longer to process.
3. Turn off web search or deep thinking for simple questions.
4. Try outside the busiest hours listed above.
5. If it never works, run a DeepSeek model locally. V4-Flash runs on a PC with about 128 GB of RAM as a 2 to 3-bit GGUF (see my DeepSeek vs Claude local section and GGUF quantization guide).

Quick Start: Call DeepSeek With the OpenAI SDK

DeepSeek’s API is OpenAI-compatible, so you can use the OpenAI Python SDK with a different base URL:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com",
    max_retries=5,  # retries 429 and 5xx with backoff
)

resp = client.chat.completions.create(
    model="deepseek-flash",   # or "deepseek-v4-pro"
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain GGUF in two sentences."},
    ],
    stream=False,
)
print(resp.choices[0].message.content)

How to Get a DeepSeek API Key and Top Up

1. Go to platform.deepseek.com and sign up with email or phone.
2. Open Top up and add balance. The API is prepaid, so calls fail with 402 Insufficient Balance once it runs out.
3. Open API keys, create a key and copy it once. Store it as an environment variable such as DEEPSEEK_API_KEY.
4. Check the Usage page after your first test calls to confirm the cache hit and peak or off-peak pricing are what you expect.

The DeepSeek chat app and website are free to use and are not billed against your API balance. The API and the app are separate products.

Flash or V4-Pro: Which One Should You Use?

A practical pattern is to send everything to Flash and escalate to V4-Pro only when Flash’s answer fails a check (bad JSON, failed unit test, low confidence). That keeps most of your bill at Flash prices.

Privacy and Where Your Data Goes

DeepSeek’s privacy policy says data is processed and stored on servers in the People’s Republic of China. Several governments and many companies have restricted DeepSeek’s app on official devices for that reason. If that matters for your project, you have two good options:

• Use the same open weights through another host. Because the models are MIT licensed, inference providers and routers such as OpenRouter offer DeepSeek models hosted in other regions, usually at a slightly higher price than DeepSeek’s own API.
• Run it yourself. V4-Flash runs locally on a PC with about 128 GB of RAM as a low-bit GGUF, so nothing leaves your machine. See DeepSeek vs Claude for the local setup.

Whatever you choose, do not send personal customer data to any API without checking your own legal requirements first.

DeepSeek vs Other APIs on Price

DeepSeek’s real edge is quality per dollar: Flash is a much larger and stronger model than the other sub-$0.30 options, and the 1M context with 384K output is unusual at this price. Full table in my AI API pricing comparison.

Frequently Asked Questions

How much does the DeepSeek API cost?

deepseek-flash costs $0.30 input and $1.20 output per million tokens at peak, and half that off-peak. deepseek-v4-pro costs $1.32 and $3.96 at peak, half off-peak. Cache hits cost a few percent of normal input.

When is the DeepSeek off-peak discount?

All hours except 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Weekends and Chinese public holidays are fully off-peak.

Is the DeepSeek API free?

No, it is prepaid. If your balance runs out you get a 402 Insufficient Balance error.

Why does DeepSeek say the server is busy?

Its servers are overloaded by traffic (503). Retry later with backoff, use off-peak hours, or use another provider or a local copy of the open weights.

Which DeepSeek model should I use?

Start with deepseek-flash. It is about four times cheaper than V4-Pro and has a 2,500-request concurrency limit. Move to V4-Pro for harder reasoning and coding.

Final Thoughts

DeepSeek remains the value pick: a 1M context, 384K output and open weights for a fraction of what the US labs charge, and half that again off-peak. The trade-offs are capability (Claude Opus 5.5 and GPT-6 Astra are clearly stronger on independent tests) and occasional capacity problems, which you can soften with retries, off-peak scheduling and a fallback model.

Compare it with the other providers in my Claude, OpenAI, Gemini and Grok pricing guides.

Sources

• DeepSeek: Models and pricing (official)
• DeepSeek: Rate limit and isolation (official)
• DeepSeek: Error codes (official)
• Tech Insider: DeepSeek V4.1-Flash-1017 MIT licence update