Canonry sentiment analysis: up to 99.4% lower estimated cost

Jev classified 2,304 AI answers for $0.37, versus an estimated $58.40 with Claude Opus 5 at the same token budget. That is a 99.4% cost reduction. Only Jev was measured; the LLM figures below are pricing estimates.

Canonry 5.30 adds sentiment scores using Jev from TypeSafe. Our open-source AEO platform now shows whether answers from ChatGPT, Gemini and Claude describe a tracked brand or property favorably, alongside whether they mention it and cite its site.

The measured run: 2,304 answers, 80 seconds, $0.37

We classified every branded answer from two sweeps of roughly 200 apartment communities across three answer engines. A sweep is one run of the portfolio's tracked questions.

Processing speed varies with answer length and account limits.

Up to 99.4% lower estimated cost than an LLM

Jev classifies each answer and returns a result with a confidence score. It does not generate a written verdict. TypeSafe charges $0.042 per million input tokens, with free output. TypeSafe pricing.

For comparison, we priced the same 8.8 million input tokens on general models, adding 250 output tokens per answer for a short JSON verdict: 576,000 output tokens in total.

Only Jev was measured. LLM costs are estimates using uncached list prices, with no batch discount or reasoning tokens. Prompts and tokenizers differ by provider, so actual input counts will differ. Cost is input millions × input rate + output millions × output rate; percentages use unrounded figures.

Prices checked September 29, 2026: DeepSeek, OpenAI, Anthropic. These costs cover sentiment classification; answer collection and hosting are separate.

About $5 a year for this portfolio

At one sweep every two weeks, this workload becomes 26 sweeps and 29,952 branded answers a year. The measured two-sweep run therefore scales by 13.

Annual figures use unrounded costs and assume unchanged volume, token use and rates. OpenAI and Anthropic offer 50% batch discounts with processing windows of up to 24 hours. For this portfolio, adding non-brand answers increases volume by about 2.4×, putting Jev at roughly $11.53 a year.

Forty seconds versus 19 hours of manual tagging

At one minute to read and label each answer, a person would spend 19.2 hours on one sweep. That is a planning assumption, not a measured human benchmark.

The per-sweep Jev figure is half the measured 80-second run. We have not timed general LLMs on this workload. People still need to review disputed verdicts and sample the results; the table compares the initial tagging step.

Accuracy: 90.3% agreement in a small human check

Two people labeled 32 answers without seeing Jev's results. We then compared their judgments with Jev.

This small check does not establish higher accuracy than an LLM. A larger human review on a separate set of answers is in progress. Sentiment is experimental, with the source quotes available for every verdict.

What the sentiment score counts

The headline is % favorable: favorable answers divided by all judged answers, shown with a 95% confidence interval and the count behind it.

Scores sit next to mention and citation coverage, per project, question and engine. The dashboard, canonry sentiment CLI, nine MCP tools and API all expose sentiment, so an agent can find poorly described properties and return the quotes behind the scores.

Turn it on for new sweeps

Sentiment is off by default. Enabling it requires both installation and project settings. Replace my-project with your project in the CLI example.

Get the source and setup instructions on Canonry's GitHub repository. Only new classifications call Jev; revisiting the evidence adds no classifier cost.