Normie ModeCan AI answer business questions from a spreadsheet?
We gave 6 AI models the same sales sheets and asked the questions a manager would ask. Every question has one right answer, so no one had to judge.
Yes, mostly. 4 of 6 models got every answer right. The rest made the most dangerous kind of mistake: a confident, plausible wrong number.
GPT-6 Sol, Gemini 3.8 Flash, Claude Opus 5.5 and Claude Sonnet 5.
When models got it wrong, they named the right customer or product but gave the wrong amount. Always sanity-check the numbers.
5 sheets × 6 questions × 6 models, run 24 Sept 2026. Total cost to run: $2.93.
Where this test counts
It's one of the tests behind the task score and picks on these pages, marked "Tests this task". The picks there combine it with the other evidence.
How each model did
Most answers right first. In the task score, models with the same number right score the same; cost and speed are shown for information.
| Model | Answers right | Cost (5 sheets) | Time per sheet |
|---|---|---|---|
| GPT-6 Sol OpenAI | 30/30 | $0.26 | 43s |
| Gemini 3.8 Flash Google DeepMind | 30/30 | $0.49 | 113s |
| Claude Opus 5.5 Anthropic | 30/30 | $0.80 | 56s |
| Claude Sonnet 5 Anthropic | 30/30 | $1.03 | 130s |
| DeepSeek V4 Pro DeepSeek | 29/30 | $0.32 | 222s |
| GPT-6 Luna OpenAI | 26/30 | $0.02 | 65s |
Results
Each model answered the same 30 questions.
Answers right
Out of 30 · Higher is better
What it cost
For all 5 sheets · Lower is better
How long it took
Average per sheet · Lower is better
Question by question
How many of the 5 sheets each model got right, for each question.
| Question | Claude Opus 5.5 | Claude Sonnet 5 | GPT-6 Sol | GPT-6 Luna | Gemini 3.8 Flash | DeepSeek V4 Pro |
|---|---|---|---|---|---|---|
| Total sales for one month | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✗ 4/5 | ✓ 5/5 | ✓ 5/5 |
| Best region last quarter | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 |
| Fastest-growing product, and by how much | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✗ 4/5 | ✓ 5/5 | ✓ 5/5 |
| Money lost to refunds and cancellations | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 |
| Biggest customer, and how much they spent | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✗ 4/5 | ✓ 5/5 | ✓ 5/5 |
| Average order size | ✓ 5/5 | ✓ 5/5 | ✓ 5/5 | ✗ 4/5 | ✓ 5/5 | ✗ 4/5 |
Where it went wrong
Every wrong answer, next to the right one.
| Model | Sheet | Question | Right answer | It said |
|---|---|---|---|---|
| GPT-6 Luna | Sheet 1 | Fastest-growing product, and by how much | Desk Lamp, 133.0% | Desk Lamp, 172.9% |
| GPT-6 Luna | Sheet 1 | Biggest customer, and how much they spent | Northwind Recruitment, £8,548.95 | Northwind Recruitment, £10,144.95 |
| GPT-6 Luna | Sheet 2 | Total sales for one month | £8,360.00 | £8,174.75 |
| GPT-6 Luna | Sheet 2 | Average order size | £557.44 | £554.59 |
| DeepSeek V4 Pro | Sheet 5 | Average order size | £765.29 | £792.57 |
See it for yourself
Sheet 1: what we asked, and what every model answered.
The questions
- What was total net revenue in March 2026?
- Which region had the highest net revenue in Q2 (April to June 2026)?
- Which product grew net revenue the most in percentage terms from Q1 (January to March) to Q2 (April to June), and by what percentage?
- How much order value was lost to refunded and cancelled orders from January to June 2026?
- Which customer had the highest net revenue from January to June 2026, and how much?
- What was the average net order value in Q1 (net revenue divided by the number of completed orders)?
The definitions we gave: an order's value is quantity × unit price minus the discount; net revenue only counts completed orders.
The sheet
123 orders, including discounts, refunds and cancellations.
Show the first 12 rows
| order id | date | customer | region | product | quantity | unit price gbp | discount pct | status |
|---|---|---|---|---|---|---|---|---|
| SO-2001 | 2026-01-01 | Juniper Health Clinic | East | Filing Cabinet | 10 | 180.00 | 5 | Completed |
| SO-2002 | 2026-01-01 | Brookside Dental | North | Monitor Arm | 11 | 65.00 | 10 | Completed |
| SO-2003 | 2026-01-02 | Elmwood Primary School | South | Desk Lamp | 2 | 38.00 | 5 | Completed |
| SO-2004 | 2026-01-03 | Pembroke Consulting | West | Desk Lamp | 2 | 38.00 | 10 | Completed |
| SO-2005 | 2026-01-04 | Copperfield Law | North | Filing Cabinet | 10 | 180.00 | 15 | Completed |
| SO-2006 | 2026-01-04 | Northwind Recruitment | West | Standing Desk | 2 | 420.00 | 0 | Completed |
| SO-2007 | 2026-01-04 | Juniper Health Clinic | East | Whiteboard | 2 | 95.00 | 0 | Completed |
| SO-2008 | 2026-01-06 | Fairhaven Estates | South | Desk Lamp | 5 | 38.00 | 0 | Completed |
| SO-2009 | 2026-01-06 | Kestrel Engineering | East | Monitor Arm | 7 | 65.00 | 0 | Refunded |
| SO-2010 | 2026-01-06 | Hollis Marketing | South | Whiteboard | 11 | 95.00 | 5 | Completed |
| SO-2011 | 2026-01-06 | Linden Travel | East | Whiteboard | 9 | 95.00 | 10 | Completed |
| SO-2012 | 2026-01-08 | Linden Travel | East | Monitor Arm | 12 | 65.00 | 5 | Refunded |
| Question | Right answer | Claude Opus 5.5 | Claude Sonnet 5 | GPT-6 Sol | GPT-6 Luna | Gemini 3.8 Flash | DeepSeek V4 Pro |
|---|---|---|---|---|---|---|---|
| Total sales for one month | £10,540.05 | ✓ £10,540.05 | ✓ £10,539.55 | ✓ £10,540.05 | ✓ £10,540.05 | ✓ £10,540.05 | ✓ £10,540.05 |
| Best region last quarter | West | ✓ West | ✓ West | ✓ West | ✓ West | ✓ West | ✓ West |
| Fastest-growing product, and by how much | Desk Lamp, 133.0% | ✓ Desk Lamp, 133.0% | ✓ Desk Lamp, 133.0% | ✓ Desk Lamp, 133.0% | ✗ Desk Lamp, 172.9% | ✓ Desk Lamp, 133.0% | ✓ Desk Lamp, 133.0% |
| Money lost to refunds and cancellations | £15,181.05 | ✓ £15,181.05 | ✓ £15,181.05 | ✓ £15,181.05 | ✓ £15,181.05 | ✓ £15,181.05 | ✓ £15,181.05 |
| Biggest customer, and how much they spent | Northwind Recruitment, £8,548.95 | ✓ Northwind Recruitment, £8,548.95 | ✓ Northwind Recruitment, £8,549.45 | ✓ Northwind Recruitment, £8,548.95 | ✗ Northwind Recruitment, £10,144.95 | ✓ Northwind Recruitment, £8,548.95 | ✓ Northwind Recruitment, £8,548.95 |
| Average order size | £807.83 | ✓ £807.83 | ✓ £807.83 | ✓ £807.83 | ✓ £807.83 | ✓ £807.83 | ✓ £807.83 |
How we tested
- Every model got identical sheets and instructions, through OpenRouter at its default settings, once per sheet.
- The sheets are made-up but realistic sales data, pasted as text. This doesn't test working with real Excel files, formulas or charts.
- Money answers count as right within 0.5% or £1; names must match the sheet.
- One run per sheet, so small differences may be luck.