Blog · The MONA effect
Same model. Same documents. 82% fewer tokens.
We put six questions about 53 long city climate contracts to Claude Opus 5 twice: once reading the files on its own, once through Harmona, with the collection indexed by MONA. Same model, same documents, same questions. With MONA, the answers came back in a fraction of the time, on a fraction of the tokens, and held their accuracy.
Tokens per question
82%fewer
489K with MONA, 2.66M without. Fewer on every single question.
Time per question
8×faster
73 seconds with MONA. Nearly ten minutes without.
Steps to an answer
57%fewer
12.5 model calls per question with MONA, 29 without.
List precision
93%
Every targeted question right. The trick question correctly refused.
What we tested
The collection is 53 Climate City Contracts from the EU Mission for climate-neutral cities: one long PDF per city, 757 MB in all. The numbers that matter sit inside tables and running text, in different units and under different headings from city to city. It is the kind of folder every team has and nobody has fully read.
We asked six questions. None of them can be answered from a file name or from general knowledge. The answer is only in the documents.
Across the collection · 3
Which cities report baseline emissions above 2,000 ktCO₂e? Which commit to a 90% or greater cut? Which use a baseline year before 2010? The answer is a list, and getting it right means checking all 53.
One document · 2
A single city's baseline figure, and the city with the highest baseline in the whole collection.
A trick · 1
The baseline of a city that has no contract in the collection. The only right answer is to say so.
Two ways to ask
Harmona with MONA
The 53 contracts uploaded to Harmona and indexed once by MONA. Claude Opus 5 answers from the indexed collection.
Claude on its own
The same model in Anthropic's own agent, reading the same 53 files directly from disk. No Harmona in between.
The results
Tokens per question
Every token the model processed to produce its answer. Lower is better.
View as table
| Question | With MONA | Without | Saving |
|---|---|---|---|
| Cities above a 2,000 ktCO₂e baseline | 1.52M | 7.68M | 80% |
| Cities committing to a 90%+ cut | 446K | 2.82M | 84% |
| Cities with a baseline year before 2010 | 569K | 2.64M | 78% |
| One city's baseline figure | 50K | 453K | 89% |
| City with the highest baseline | 263K | 2.24M | 88% |
| A city with no contract | 79K | 160K | 51% |
Time and steps per question
Across all six questions. Lower is better.
Median seconds per question
Model calls per question
View as table
| System | Median seconds | Model calls | Mean tokens |
|---|---|---|---|
| Harmona with MONA | 73 | 12.5 | 488,534 |
| Claude on its own | 575 | 29.2 | 2,664,136 |
And the answers held
Both got every targeted question right: the single city's figure and the highest baseline in the collection. Both recognised the trick and said the document was not there. On the three collection-wide questions, where the answer is a list of cities, MONA's lists were 93% precise and landed within a few points of Claude reading every file itself.
On the hardest of them, three qualifying cities hidden in 53 documents, the two gave the identical answer. Harmona reached it on a sixth of the tokens and in a fifth of the time.
Why MONA changes the numbers
Without MONA, the model opens the files afresh for every question and carries everything it has read through every step. The biggest sweep took it 74 steps and 7.7 million tokens.
The saving showed on every one of the six questions, from the single-city lookup to the largest sweep: between half and a ninth of the tokens, 82% fewer on average.
What it means for your team
Pay for the questions, not the folder.
MONA indexes your documents once into a corporate knowledge base. Every question after that costs a fraction of reading the files again.
Answers in a minute, not ten.
A median of 73 seconds per question. The slowest question without MONA took 24 minutes.
Lists you can use.
93% precision means the lists MONA returns need little correction before they go into a report.
Honest when the document is missing.
Asked about a city with no contract, Harmona said so and pointed to where the city was only mentioned in passing.
See the MONA effect on your own documents
Connect your documents, let MONA index them into your corporate knowledge base, and ask the questions your team has been putting off.
Talk to usAbout this test
- Claude Opus 5 on both sides. What differs is whether MONA has indexed the documents first.
- The files were uploaded to Harmona and indexed by MONA. Anthropic's own agent read the same files directly from disk.
- Answers were scored against an answer key in which every value is backed by a verbatim quote from the source document. On the three list questions, the accuracy score (F1, which rewards finding every right city and penalises naming wrong ones) was 84% for Harmona and 88% for Claude. A four-point difference is within the variation we see when the same question is asked again.