Workflows / paytm-settlement-queries · Compare · compared 20 Oct, 3:30 pm by ops.member

Compare models

Done: dvar:deepseek-v4-flash-0731 passes · 89% cheaper
Request a fine-tune

Gold v1Rubric v1quality bar ≥ 84%p99 ≤ 15,000 ms 216 gold rows replayed.

Results. One row per model, scored on your gold set. Candidates replayed on 160 gold rows; today's model from its benchmark replies on the same gold rows.
ModelQualityp99 latencyCost per requestvs todayResult
Gemini 3.8 Flash Stand-in for today's modelopenrouter/google/gemini-3.8-flash85.0%36,399 ms₹0.4219—Today
DeepSeek V4 Flashopenai_compatible/dvar:deepseek-v4-flash-073188.8%12,419 ms₹0.0468−89%Recommended
openai_compatible/dvar:glm-5.3-flash85.0%10,491 ms₹0.0579−86%Fails 1 check
openai_compatible/dvar:gemma-4-31b-it83.1%27,622 ms₹0.0718−83%Fails 3 checks
openrouter/anthropic/claude-sonnet-5.587.5%19,510 ms₹0.6816+62%Fails 2 checks
Post-trained Dvar model: next step, measured after training on the assistant's outcomes. No numbers yet. It goes through this same gate: Fine-tune.

DeepSeek V4 Flash passes every check: 3.8 points higher than today's model at 89% less cost.

Prices are per request, at each provider's list price.

Judge: 7 of 7 checks trusted.

Score by topic. Score for each topic, today against each candidate.Scale 0–1minimum per topicbelow minimum
TopicRowsToday · Gemini 3.8 FlashCandidate · DeepSeek V4 FlashCandidate · GLM 5.3 FlashCandidate · Gemma 4 31BCandidate · Claude Sonnet 5.5
Settlement status48
0.92
0.92
0.00
0.85
−0.07
0.92
0.00
0.94
+0.02
Delayed, on hold or failed26
0.92
0.88
−0.04
0.85
−0.07
0.92
0.00
1.00
+0.08
UTR and proof19
0.89
0.89
0.00
0.95
+0.06
0.79
−0.10
0.95
+0.06
Account and plan changes12
0.83
0.83
0.00
0.83
0.00
0.83
0.00
0.67
−0.16
Amount and deductions36
0.75
0.92
+0.17
0.83
+0.08
0.75
0.00
0.75
0.00
Payouts13
0.69
0.69
0.00
0.85
+0.16
0.77
+0.08
0.85
+0.16
Instant settlement6
0.83
1.00
+0.17
0.67
−0.16
0.50
−0.33
0.83
0.00
All topics, weighted by traffic160
0.85
0.89
+0.04
0.85
0.00
0.83
−0.02
0.88
+0.03
——