Nova, our internal customer: one month on Fronset

First, a disclosure. The customer here is our internal customer, Nova (https://nova.kapualabs.com), a content and financial-analysis publication run by the same team that builds Fronset. Its pipeline has called the Fronset API like any account since 15 July 2026. It is the account with the longest history, so it is the one we can measure. It is not a typical customer: it runs a lot of small, cheap steps.

The question

A team with a pipeline like this, and no time to test every model on every step, sends every step to a model it trusts to do reasonably well. A typical choice is a mid-tier model, Claude Sonnet 5. A cautious team picks the top model, Claude Opus 5. Fronset does it differently: you name a task, and each call runs on the cheapest model that graded outputs of that task show is good enough.

So we priced one month of Nova’s traffic three ways.

September 2026

The calls cover 25 kinds of step: claim extraction and refinement, language identification, summarisation, translation, relevance scoring, analytical writing and smaller ones. 76 calls on image tasks and on our own provider keys are left out of all three columns.

How we priced Sonnet 5 and Opus 5

We did not just multiply our token counts by Anthropic’s price list. In September both models ran on some of these same requests, as alternatives Fronset tried while measuring the tasks: 223 attempts by Sonnet 5 and 275 by Opus 5. From those attempts we took what each model really does with these inputs:

The input row is Anthropic’s own counting: on the same requests and the same prompts, every other provider we measured counts 0.86–1.05 times the tokens, and Anthropic’s newer models count more. A team calling Anthropic directly pays for those tokens too.

Each call was then priced at Anthropic’s list price at the time, with its batch discount where the call ran in batch, and multiplied by what we really pay Anthropic against that list.

Quality

Fronset’s judges grade answers from many models on every task. Weighted by how many calls each step made, the answers Fronset served scored 8.45 out of 10; Sonnet 5 scores 8.52 on the same tasks and Opus 5 8.66. Overall, Fronset and Sonnet 5 are within the measurement’s noise of each other (about ±0.3), and Opus 5 is a little better.

Step by step it differs:

• On claim refinement, our second-largest step, Sonnet 5 was clearly better: 8.33 against Fronset’s 7.79.
• On summarisation and relevance scoring Fronset’s choices were clearly better: 8.93 against 8.11, and 7.07 against 5.10.

So: about the same quality as Sonnet 5 for about a quarter of the price, and a little under Opus 5 for about a ninth.

What this does not show

• One month and one account. Our pipeline runs many small steps. A pipeline of long, hard steps would spend more of its bill on the models and less on our fee.
• The direct prices rest on a few hundred attempts per model. For Opus 5 the failure cost is driven by two long-writing steps where most of its attempts failed. Counting no failures at all, it would still cost about $5,400.
• The quality scores are per-task averages for each model, read in October, not scores of these exact answers.
• September was not invoiced. It came before we started charging real money (7 October); the Fronset figure is what our meter recorded for every call, at the rates and fees a customer would have paid.
• Both models have since been superseded. Claude Sonnet 5.5 replaced Sonnet 5 on 12 October. They were the models on sale in September.