Case study 02 · LLM cost and routing

LLM cost-per-build routing

Under 700 credits per buildGLM chairs · Opus writes$99 flat → usage pricing

An AI app builder that asks several models to work through every build can get expensive quickly. This is how Quantlys Conclave splits the work between models by job, the budget each build is held to, and the same thinking applied to a voice product's pricing.

01 · The problem

A build is many model calls

In Conclave, a build isn't one prompt. A crew of agents (PM, Architect, Developer and QA) moves the work through four stages: Draft, Peer-rank, Arbiter and Ship. Several models draft a plan, rank each other's drafts, an Arbiter merges the best ideas into one plan, and the crew builds and checks the app. Every stage spends tokens, and every build is measured in credits.

If every step uses the strongest model, quality is high and so is the bill. If every step uses the cheapest one, files break and builds have to be re-run, which costs more in the end. So the useful question isn't “which model is best?” but “which model is best for this step, at this price?”

02 · The approach

Budget first, then route by job

A number per build. The target is under 700 credits per build, and under 900 for CRM-class apps. A budget turns “is this too expensive?” into a check instead of a debate.

Three criteria, not one. Models are chosen on price, quality and data retention. What a provider keeps, and for how long, is part of the cost.

Route by job. On November 4, 2026, the split became: GLM chairs the discussion, and Opus writes the files. The discussion is long back-and-forth reasoning, where a lower-cost model holds up well. Writing files is where a mistake means a re-run, so that step goes to the model that's most reliable at it. A separate trial is checking whether more of the savings can be recovered without giving up that reliability.

03 · What happened

Targets, and a pricing rethink

The split is recent, so 700 and 900 are targets, not results. I won't publish an average until enough builds are behind it. What changed right away was the conversation: once each build had a budget, choosing a model became a trade-off with a number attached, not a matter of taste.

I'd learned the same lesson earlier on Tablerene, an AI phone host for restaurants built with Conclave. It started at a flat $99. After I mapped what Twilio actually charges for calls and texts, pricing moved to usage: $29 a month per location, $0.18 per answered minute and $0.03 per text. Billing counts only answered seconds, each restaurant gets one number, spend caps are on, and a card has to be on file before a phone number is bought.

What hasn't happened yet: paying customers haven't tested Tablerene's pricing. It works on paper. The market hasn't voted.

04 · Lessons

What the numbers taught me

  • Budget the unit the customer sees, such as a build or a call, not just the monthly API bill.
  • Route by step. Use cheaper models where mistakes are easy to catch and reliable ones where mistakes are expensive to redo.
  • Data retention is part of the price.
  • Let the cost driver set the price. Flat pricing on a product with usage-based costs moves the risk onto you.
  • Targets aren't results. Say which one you have.
05 · For another team

Before you pick a model

Write down what one unit of work may cost before you choose a model, and check it again whenever a model or a price changes. Split your pipeline into steps and ask, for each one, what a mistake there costs. Expect the answer to be a mix of models, and expect the mix to change, so keep routing easy to adjust.

To see how Quantlys compares with other AI app builders, read Best Lovable alternatives on quantlys.ai.

About the author

Santosh Adari

Santosh Adari is Associate Director, IT Data Engineering at Gilead Sciences and the founder of Quantlys, which he builds on nights and weekends from Apex, North Carolina. Quantlys is a native app for Apple Silicon Macs; inside it, Conclave plans an app with a council of AI models, then builds it and deploys it to your own GitHub, Vercel and Supabase accounts.

About Santosh Adari · All case studies · Built with Quantlys · LinkedIn