DeepSeek V3.2 API in India

DeepSeek V3.2 is the reasoning-first model in the DeepSeek family — built for problems that need many steps of thinking, not quick replies. On unoblox it sits behind the same OpenAI-compatible endpoint as every other model, billed at ₹26.21 per million input tokens and ₹38.30 per million output tokens on your monthly GST invoice.

When to reach for V3.2

V3.2 is deliberately slower and pricier than the Flash tier because it spends more tokens thinking through a problem. Reach for it when the task actually needs that: theorem-style proofs, multi-step debugging, scientific derivations, or reasoning across a long document where a shallow answer isn't good enough. For everyday chat, summarization, or classification, DeepSeek V4 Flash is the better — and cheaper — fit.

V3.2 pricing versus the rest of the catalog

Model	Input (₹/1M)	Output (₹/1M)	Best for
DeepSeek V3.2	₹26.21	₹38.30	Deep reasoning, proofs, research
DeepSeek V4 Flash	₹9.07	₹18.14	High-volume chat, RAG
GPT-4o	₹252	₹1008	Multimodal production apps
Claude Opus	₹504	₹2520	Frontier general reasoning

Every rate above is live on /models and billed straight in rupees — no international card, nothing to reconcile against a foreign invoice.

How to call V3.2

1. Create an account → https://unoblox.ai/sign-in and generate a key (ub-gw-…).
2. Point your SDK at unoblox → base_url = https://api.unoblox.ai/v1.
3. Set the model → deepseek-ai/deepseek-v3.2, with a generous max_tokens since reasoning consumes output tokens.

from openai import OpenAI

client = OpenAI(api_key="ub-gw-...", base_url="https://api.unoblox.ai/v1")
completion = client.chat.completions.create(
    model="deepseek-ai/deepseek-v3.2",
    messages=[{"role": "user", "content": "Walk through this proof step by step..."}],
    max_tokens=4000,
)

Where V3.2 earns its keep

• Research and analysis — synthesizing papers, spotting contradictions, building hypotheses.
• Hard math and logic — olympiad-style problems, multi-step algebra, formal proofs.
• Code review for correctness — tracing subtle bugs through cryptography or concurrent code.
• Scientific computation — multi-step derivations where each step has to be right before the next.

Frequently asked questions

Is V3.2 overkill for a simple chatbot? Usually, yes. Use DeepSeek V4 Flash instead — you'll save roughly ₹17 per million input tokens and get a snappier response for tasks that don't need deep reasoning.

How do I control how much V3.2 "thinks"? Through max_tokens. Reasoning tokens count as output, so set a generous ceiling (3000+) for genuinely hard problems and a smaller one when you want a tighter answer.

Can I stream V3.2's output? Yes — pass stream=true and tokens arrive as the model reasons, which is useful for showing progress on long-running requests.

Does unoblox charge extra for longer reasoning chains? No. You pay standard per-token output pricing regardless of how many internal steps the model takes — there's no separate "thinking" fee.

What if I need something between V3.2 and V4 Flash? Check /models for the full DeepSeek line-up and current rates — new tiers are added as they go live.

Get started in rupees → https://unoblox.ai/sign-in