What 2.5 lakh filings a month cost
Same pipeline, same pages, same checks. Only the model changes. Token counts come from metered runs on 147 real filings. Change any number below and everything updates.
| Model | Per filing | Per 1,000 | Per month | Per year |
|---|
GLM 5.3 Flash prices are Dvar's rates. Opus 5.5 is $4 in and $20 out per million tokens at list price. Its row applies those prices to GLM's token counts until the Opus run below fills in its own; Opus counts tokens differently and spends some output on thinking, so treat that row as a floor. The mixed row runs GLM on everything and also sends the shadow sample to Opus.
How much accuracy does GLM give up against Opus?
Two ways to measure it, and they answer different questions. The first is a one-time test on filings a person has already labelled. The second keeps measuring on live traffic for a small, fixed cost.
1. The same 130 labelled filings, both models
| Model | Correct | Published wrong | Held | Error rate, 95% bound | ₹ per filing |
|---|
"Error rate, 95% bound" is the highest error rate that still fits the result. With 0 misses in 130, the true rate is most likely under 2.3%, not zero. Telling apart two models that are both under 2% needs more filings than this set has, which is what the second method is for.
2. Keep measuring on live filings: a shadow sample
GLM answers every filing as usual. A random slice also goes to Opus in the background. Where the two agree, nothing happens. Where they disagree, a person opens the PDF and says who was right. That gives the accuracy gap on real traffic, month by month, and a person only reads the disagreements.