How Indian inference pricing works
Indian inference pricing is usually quoted in INR per 1 million tokens. Providers separate input, cached input, and output because the cost of reading a long prompt is different from the cost of generating a long answer. Input covers the prompt and any retrieved context. Cached input covers repeated prompt prefixes that the provider can reuse. Output covers the generated answer, including reasoning tokens on reasoning models.
The spread is wide. Small open models such as Gemma 4 E4B and Qwen3.5 9B start at ₹3 per 1M input tokens on Krutrim. Frontier models such as DeepSeek V4.1 Flash and GLM 5.3 Flash start lower on input but charge more for output. A cheap input rate is not a cheap workload if the answer is long.
Most Indian providers quote list prices in INR and exclude 18% GST. Some also publish a USD reference, which moves with the exchange rate. Cache rates are often 10% to 50% of the input rate, but not every provider publishes one. A dash means the provider does not list a cache price, not that caching is free.
For the full price list, see India inference pricing.