KreditBeeDvar

Inference cost

What the model calls cost at lending volume

Unit costs come from the tokens each model used on the test sets and its list price. Volumes are yours to change.

Internal benchmark on synthetic data · run run-20261022172936-test · 2026-10-22 · 3 models, same instruction

WorkflowMonthly volume (yours to change)Dvar small modeldvar:glm-5.3-flashLow-cost frontieropenai/gpt-6-lunaMid-tier frontieranthropic/claude-sonnet-5.5
Post-call dispositionsmeasured tokens, collections runcalls a month₹0.33 lakh$0.26 per 1,000₹0.14 lakh$0.11 per 1,000₹4.21 lakh$3.34 per 1,000
Bank statement labelsmeasured tokens, underwriting runstatements a month₹0.54 lakh$1.07 per 1,000₹0.22 lakh$0.45 per 1,000₹4.84 lakh$9.60 per 1,000
Calling-agent replies (LLM only)assumed 1,400 in / 60 out tokens a turn, 7 turns a callcalls a month₹1.72 lakh$1.37 per 1,000₹1.50 lakh$1.19 per 1,000₹29.99 lakh$23.80 per 1,000
Model cost a month₹2.59 lakh$3,081₹1.87 lakh$2,222₹39.03 lakh$46,469

At these volumes the Dvar small model costs 15× less than the mid-tier frontier model. The low-cost frontier model costs 1.4× less than the Dvar small model. Price alone does not decide; the critical failures, latency and where the data is served do, and the other pages measure those.