Gemma API Pricing in India
Gemma models run on Lumino's OpenAI-compatible chat completions API with streaming, per-token billing in INR, and usage visibility. Gemma is Google's open-weight family, and the main reason teams pick it over a proprietary frontier model is cost: Gemma 4 26B delivers usable quality at a fraction of the price of a hosted frontier tier while keeping a 256,000-token context window.
Gemma models and pricing
| Model | Input / 1M | Output / 1M | Context window | Max output | Status |
|---|---|---|---|---|---|
| Gemma 4 26B | ₹10 | ₹50 | 256,000 tokens | 8,192 tokens | Available |
| Qwen3 Embedding 0.6B | ₹20 | — | 4,096 tokens | — | Embedding, currently disabled |
Gemma 4 26B costs ₹10 per million input tokens and ₹50 per million output tokens. For comparison, a hosted frontier coding tier on the same platform runs ₹200 to ₹380 per million input. For chat, summarisation, extraction, and Q&A workloads the quality gap rarely justifies that multiple.
What Gemma costs in practice
Assume a summarisation workload: 3,000 input tokens of document text producing 600 output tokens of summary.
- One call costs about ₹0.06 (₹10 × 3,000 / 1M + ₹50 × 600 / 1M).
- At 1,000 calls per day that is roughly ₹1,800 per month.
- Running the same 1,000 calls per day on a frontier tier at ₹200/₹390 would cost about ₹23,400 per month.
High-volume summarisation, classification, and customer-support drafting are exactly where this difference decides whether a feature is viable.
Where Gemma fits and where it does not
| Workload | Verdict |
|---|---|
| Document summarisation | Strong fit. Long context, cheap tokens, predictable output length. |
| Classification and extraction | Strong fit. Output is short, so the low output rate dominates. |
| Customer support drafting | Good fit. Style transfer works well at this size. |
| Multilingual chat | Good fit, particularly across Indian and European languages. |
| Hard multi-step reasoning | Weak fit. Step up to a reasoning-focused tier. |
| Long autonomous agent loops | Poor fit. Compounding output errors make a stronger model cheaper overall. |
Gemma 4 26B architecture notes
Gemma 4 26B is a mixture-of-experts model, so it activates only part of its parameters per token. That is why it behaves like a much larger model on quality while costing like a mid-sized one. Two practical consequences:
- Latency stays low because only the active expert path runs per token.
- It does not support reasoning mode, so it will not emit chain-of-thought. If you need explicit reasoning traces, pick a reasoning tier instead.
Calling the Gemma API
curl https://api.luminoai.co.in/api/v1/chat/completions \
-H "Authorization: Bearer nm-your-key" \
-H "Content-Type: application/json" \
-d '{
"model": "gemma-4-26b",
"messages": [{"role": "user", "content": "Summarise this contract in 5 bullet points."}],
"max_tokens": 1024,
"stream": true
}'Model ids and availability should always be confirmed in the live model catalog before a production rollout.
Gemma API or a rented GPU?
| Decision point | Practical rule |
|---|---|
| Use the hosted Gemma API when | You want chat or summarisation features without running a GPU server, and per-token cost is the binding constraint. |
| Use a rented GPU when | You need custom weights, fine-tuned checkpoints, or full runtime control over serving flags. |
| Plan cost by | Input tokens, output length, retries, regenerate behaviour, and traffic bursts. |