Ask your spend · For a finance admin or a CEO
Plain-English questions, answered from the ledger
- The problem today
- A simple spend question means an export, a spreadsheet and a wait.
- What it does
- Answers in seconds. The model only writes a small query; code runs it on the ledger, so every figure is computed from the books.
- What you see
- The question, the query the model wrote, and the answer and chart computed from the ledger.
Internal benchmark on synthetic data · run run-20261022172600-test · spend questions in the test split
Fastest on this run: Mid-tier frontier. Most answers right: Mid-tier frontier. A wrong query gives a wrong total, but never a figure the ledger does not hold.
Show figures for each modelHide figures
Dvar small model
- Answer right to the cent
- 27 of 30
- Median time to the query
- 1.57 s
- Model cost per question
- $0.00020
Low-cost frontier
- Answer right to the cent
- 27 of 30
- Median time to the query
- 714 ms
- Model cost per question
- $0.00011
Mid-tier frontier
- Answer right to the cent
- 28 of 30
- Median time to the query
- 533 ms
- Model cost per question
- $0.0028
| Model | Answer right to the cent | Median time to the query | Model cost per question |
|---|---|---|---|
| Dvar small model | 27 of 30 | 1.57 s | $0.00020 |
| Low-cost frontier | 27 of 30 | 714 ms | $0.00011 |
| Mid-tier frontier | 28 of 30 | 533 ms | $0.0028 |
Fastest on this run: Mid-tier frontier. Most answers right: Mid-tier frontier. A wrong query gives a wrong total, but never a figure the ledger does not hold.
Internal benchmark on synthetic data · run run-20261022172600-test · 2026-10-22 · 3 models, same instruction
Ask about spend from April to October in plain words. Pick a question to start.
Behind the answer
- Model
- dvar:glm-5.3-flash
- Time
- —
- Cost per question
- —
- Scorer
- Press the button above
The scorer runs the model's query and the answer key's query over the same ledger and compares the results to the cent. A query that does not fit the shape is refused and nothing is shown.
Run run-20261022172600-test, 2026-10-22. Same instruction for every model, temperature 0. Cost is from tokens and list prices.
Why this is safe to put in front of a finance admin
- The model writes a small JSON query, never a figure. Code checks its shape and runs it.
- A query with an unknown field, a bad period or anything like SQL is refused before it runs.
- The step is small (one short query per question), so it can be measured per model for accuracy, time and cost: see the table at the top of this page.