Batch AI API pricing in India

"Batch" usually implies a discounted, asynchronous pricing tier on other platforms — unoblox doesn't currently publish a separate discounted batch tier, so the honest answer for pricing a large offline job is simpler: every request costs the same per-token ₹ rate whether you send it as part of an interactive chat or as the 10,000th row of an overnight script. That makes sizing a batch job pure arithmetic once you know your token counts.

The assumption this example uses

Assume a one-off backlog of 10,000 documents, each sending about 800 input tokens (source text plus instructions) and generating about 150 output tokens (a summary or label) — 8.0M input and 1.5M output tokens for the whole job. Swap in your own document count and token size before trusting a total.

Cost for the whole backlog, by model

Model id	₹ / 1M input	₹ / 1M output	₹ for this whole job
deepseek-ai/deepseek-v4-flash	₹9.07	₹18.14	₹99.77
qwen/qwen3-235b-a22b-instruct-2507	₹9.07	₹55.44	₹155.72
meta-llama/llama-4-scout-17b-16e-instruct	₹10.08	₹30.24	₹126.00
openai/gpt-5-mini	₹25.20	₹201.60	₹504.00
openai/gpt-4.1	₹201.60	₹806.40	₹2,822.40

Running this overnight instead of spread across business hours doesn't change the total — unoblox meters tokens, not wall-clock time or request scheduling, so "batch" here describes how you schedule the work against the same standard endpoint rather than a separate price list. Claude and other models not shown above should be priced from their live /models rate before you commit a large job to them.

Why no discount claim

Some providers publish a lower per-token rate for asynchronous, delayed-response batch jobs. unoblox doesn't currently advertise one, so this page doesn't assume a discount exists — if a dedicated batch/async tier is introduced, it will show up directly as a different rate on the affected model's /models page, not as a multiplier you have to remember.

Keeping a large job predictable

Because the total is just tokens times rate, the practical way to control a batch job's cost is the same as any other workload here: trim the per-document prompt to what the task actually needs, cap output length (a one-line label costs far less than a paragraph), and test the per-document token count on a small sample before running the full backlog.

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":"Summarise this document in one sentence."}]}'

Frequently asked questions

Does unoblox have a discounted batch API tier? Not currently — every request is billed at the same standard ₹ per-token rate shown on the model's page, regardless of how it's scheduled.

Does running jobs overnight cost less? No — cost is metered by tokens, not by time of day or how long a job takes to finish.

How do I estimate a large job before running it? Run a small sample, note the average input and output tokens per item from your usage dashboard, then multiply by your full item count and the model's listed ₹ rate.

Is there a free model for testing a batch script? Yes — qwen/qwen3-1.7b is priced at ₹0, useful for validating a pipeline's logic before switching to a paid model for the full run.

What's the single biggest lever on a batch job's total cost? Output length per item, in most document-processing jobs — a short label or one-line summary costs a fraction of a full paragraph at the same input size.

How is the invoice structured for a large one-off job? It's folded into the same monthly GST invoice as the rest of your usage — in rupees, from an Indian entity, input tax credit claimable.

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