Qwen API Pricing in India
Qwen models are served through Lumino's OpenAI-compatible chat completions API. Qwen is the workhorse for coding assistants, agent workflows, and multilingual product features in India, largely because the smaller tiers stay cheap per token while still handling a 256K context window. Pricing below is INR per million tokens.
Qwen models and pricing
| Model | Input / 1M | Output / 1M | Context window | Max output | Best for |
|---|---|---|---|---|---|
| Qwen3.5 397B A17B | ₹70 | ₹450 | 262,144 | 81,920 | Maximum capability, currently disabled |
| Qwen3.7 Max | ₹380 | ₹1,130 | 256,000 | 32,768 | Hard agentic coding, long autonomous loops |
| Qwen3.8 Max | ₹260 | ₹760 | 250,000 | 32,768 | Production coding default |
| Qwen3.8 27B | ₹30 | ₹290 | 256,000 | 16,384 | Mid-tier quality, full context |
| Qwen3.6 35B-A3B | ₹20 | ₹140 | 131,072 | 16,384 | Budget agentic coding |
| Qwen3.8 Flash | ₹20 | ₹60 | 976,000 | 16,384 | High-volume, near-1M context |
Within the Qwen lineup the spread is enormous: Qwen3.7 Max output costs about 19x Qwen3.8 Flash output. Picking the tier per task rather than per project is the single biggest saving available.
What a coding-agent turn costs
A realistic coding agent sends 4,000 input tokens of file context and expects 1,200 output tokens of code plus explanation.
- Qwen3.7 Max: about ₹2.88 per turn. At 300 turns per day, roughly ₹25,900 per month.
- Qwen3.8 Max: about ₹1.95 per turn, roughly ₹17,600 per month.
- Qwen3.8 Flash: about ₹0.15 per turn, roughly ₹1,360 per month.
- Qwen3.6 35B-A3B: about ₹0.25 per turn, roughly ₹2,240 per month.
A two-tier setup — Flash for reading, planning, and file search, Max only for generating the final patch — usually lands between the Flash and Max totals while keeping most of the quality.
Why Qwen output length is the main cost driver
On every Max tier output is priced about 2.9x input. Coding models also tend to restate the problem before answering, which inflates output tokens on every single call.
- Set max_tokens to a realistic ceiling. An agent that is allowed 32K will sometimes use 30K on a task that needed 2K.
- Truncate tool output before it re-enters history. A
catof a large file can dominate the next three requests. - Cap agent loop iterations. A retry loop that re-sends full context bills the context again each pass.
Calling the Qwen API
curl https://api.luminoai.co.in/api/v1/chat/completions \
-H "Authorization: Bearer nm-your-key" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-8-max",
"messages": [{"role": "user", "content": "Refactor this module and explain the change."}],
"max_tokens": 4096,
"stream": true
}'One key covers every Qwen model, including qwen3-8-flash for high-volume traffic and qwen3-7-max for the hardest reasoning.
Qwen API or a rented GPU?
| Decision point | Practical rule |
|---|---|
| Use the hosted Qwen API when | Traffic is bursty, you need API-first coding behaviour, and you do not want to run inference servers. |
| Consider a rented GPU when | You need a self-hosted checkpoint, custom serving flags, or sustained high-utilisation inference. |
| Validate before launch by | Testing representative prompts, max output length, retry behaviour, and rate limit headroom on the exact tier you plan to ship. |