DeepSeek API Pricing in India

DeepSeek models and pricing

The gap between V4 Pro and V4 Flash is about 13x on input and 13x on output. Model choice is usually the single largest cost lever on a DeepSeek workload, larger than any prompt optimisation.

What a DeepSeek request costs in practice

Take a typical coding-agent turn: 2,000 input tokens of repository context plus 800 output tokens of generated code and explanation.

• On DeepSeek V4 Pro that single turn costs roughly ₹0.71 (₹200 × 2,000 / 1M + ₹390 × 800 / 1M).
• On DeepSeek V4 Flash the same turn costs about ₹0.04.
• At 500 such turns per day, Pro runs about ₹10,650 per month and Flash about ₹600.

This is why short agent loops are expensive on Pro and cheap on Flash, and why routing by task difficulty beats picking one model for everything.

Why DeepSeek output length drives your bill

DeepSeek output is priced about 2x input on Pro and 3x on Flash. Reasoning models also generate hidden reasoning tokens before the visible answer. Three patterns dominate real-world DeepSeek spend:

• Unbounded max_tokens. The model uses whatever ceiling you allow. Cap it to what the task actually needs.
• Retry loops. A failed tool call that re-sends full history bills the whole history again on every attempt.
• User regenerate. Every "try again" is a fresh billed request. Rate-limit the regenerate button, not just the API.

Trimming input context and setting a realistic max_tokens typically cuts a DeepSeek bill more than switching models.

Calling the DeepSeek API

One key covers every DeepSeek model. Send a standard OpenAI chat completions request and change only the model id.

curl https://api.luminoai.co.in/api/v1/chat/completions \
  -H "Authorization: Bearer nm-your-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [{"role": "user", "content": "Review this diff for race conditions."}],
    "max_tokens": 2048,
    "stream": true
  }'

Swap deepseek-v4-pro for deepseek-v4-flash or deepseek-v4-1-flash to change tier without touching any other code.

DeepSeek API or a rented GPU?

Related pages