AI agent running cost in India

An AI agent — a loop that reads context, calls tools and decides what to do next — burns far more tokens per useful action than a single chat reply, because the model re-reads the running history, the tool schemas and its own prior steps on every turn. Since unoblox bills every model by the token at a listed ₹ rate, an agent's running cost in India is arithmetic, not guesswork, once you know roughly how many tokens each run consumes.

The assumption this example uses

Assume an agent handles 1,000 runs a day, and each run sends about 3,000 input tokens (conversation history, tool schemas, retrieved context) and generates about 1,500 output tokens (reasoning, tool calls, final reply) — roughly a 2:1 input:output ratio, for 3.0M input and 1.5M output tokens across the day. Your own agent's real ratio depends on how much history you replay each turn and how verbose its reasoning is — treat this as a worked example, not a universal number.

What that costs by model

Model id	₹ / 1M input	₹ / 1M output	₹ / day at this volume
deepseek-ai/deepseek-v4-flash	₹9.07	₹18.14	₹54.42
qwen/qwen3-235b-a22b-instruct-2507	₹9.07	₹55.44	₹110.37
meta-llama/llama-4-scout-17b-16e-instruct	₹10.08	₹30.24	₹75.60
openai/gpt-5-mini	₹25.20	₹201.60	₹378.00
openai/gpt-5	₹126.00	₹1,008.00	₹1,890.00

Claude Sonnet (anthropic/claude-sonnet-4-5/-4-6/-5, ₹201.60 / ₹1,008.00 per 1M tokens) would run ₹2,116.80/day at the same volume. qwen/qwen3-1.7b is priced at ₹0, which makes it a reasonable choice for developing and testing the agent loop itself before switching to a paid model for production runs. For any model not shown here — o3, o4-mini, Claude Opus, Nemotron and the rest of the catalog — read the live ₹ rate on /models before sizing a budget around it.

Why the spread is this wide

At this example's volume, the cheapest model in the table and Claude Sonnet differ by roughly ₹2,062.38 a day — not because one model is "better" in a way this page can quantify, but because output tokens are the expensive side of every listed price, and agents generate a lot of them (reasoning text, tool-call arguments, retries). A model priced low per output token is worth evaluating first for high-volume agent workloads; save the priciest models for steps that clearly need them.

Levers that actually reduce the bill

Because every model here is billed per token, the direct way to lower an agent's running cost is to send fewer tokens per turn: trim tool schemas to only what the current step needs, summarise older conversation turns instead of replaying them verbatim, and cap output length for steps that don't need long reasoning. None of that changes the ₹ rate — it changes how many tokens you're paying that rate on.

A minimal request

curl https://api.unoblox.ai/v1/chat/completions \
  -H "Authorization: Bearer ub-gw-xxxxxxxxxxxxxxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-ai/deepseek-v4-flash","messages":[{"role":"user","content":"Plan the next tool call."}]}'

Frequently asked questions

Does unoblox charge extra for tool/function calling? No — tool-call tokens (the schema you send and the arguments the model returns) are billed as ordinary input/output tokens at the model's listed rate.

Which model should I start with for a new agent? A cheaper model such as DeepSeek V4 Flash or Qwen3 235B-A22B for the bulk of turns, reserving a stronger model for steps that clearly need it — validate quality on your own workload rather than assuming from price alone.

Is there a free option while I build the agent loop? Yes — qwen/qwen3-1.7b is listed at ₹0, useful before you point production runs at a paid model.

How accurate is the token estimate for my own agent? Only as accurate as your assumptions about history length and output verbosity — check actual token counts for your workload in the account usage dashboard rather than relying solely on a worked example.

Do longer agent chains always cost proportionally more? Yes, in tokens — more turns means more input replay and more output, so the total scales with however many turns your agent actually takes, not just the per-run estimate above.

What if my agent calls a model that isn't in the ₹ price list? Check its live rate on /models before estimating cost — don't assume it's priced like a neighbouring model.

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