Best AI model for accounting firms, tested on 23 September 2026
The tasks in the AIGROW small-business benchmark that an accounting firm would hand to AI, and how each of 36 models did on every one, marked pass or fail against written checks.
A pilot, 1 run a model on each task. This test ran on 23 September 2026. Since then makers have replaced 16 of its 36 models with newer ones, and it did not include 16 current models on our list. The LLM leaderboard ranks current models on measured results.
- Qwen3.8 Flash
- 6 of 6
- Kimi K3
- 6 of 6
- Gemini 3.1 Pro (preview)
- 6 of 6
- GLM-5.3-Flash
- 6 of 6
- DeepSeek V4.1 Flash
- 6 of 6
- GLM-5.3
- 6 of 6
The AIGROW small-business benchmark has 6 tasks an accounting firm would hand to AI: categorising an Italian bank statement, categorising bank transactions, extracting invoice data, extracting Italian supplier invoices, calculating Italian professional invoices and calculating overtime pay. On 23 September 2026, 15 models passed all 6; of those, gpt-oss-120b gave the cheapest passes across the benchmark, $0.422 per thousand.
Every model on these tasks
Most passes first; a tie goes to the better record across all 20 tasks. The cost of 1,000 passes is the whole benchmark’s, failed runs included.
| Model | Categorising an Italian bank statement | Categorising bank transactions | Extracting invoice data | Extracting Italian supplier invoices | Calculating Italian professional invoices | Calculating overtime pay | Passed | 1,000 passes |
|---|---|---|---|---|---|---|---|---|
| Qwen3.8 Flash | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $1.09 |
| Kimi K3Replaced by Kimi K3.1 Pro | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $18.44 |
| Gemini 3.1 Pro (preview)Replaced by Gemini 4 Argon | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $29.67 |
| GLM-5.3-Flash | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $0.831 |
| DeepSeek V4.1 Flash | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $1.05 |
| GLM-5.3 | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $7.02 |
| Grok 4.7 | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $9.41 |
| Qwen3.7 Plus | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $3.58 |
| Claude Sonnet 5Replaced by Claude Sonnet 5.5 | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $6.70 |
| GPT-5.6 SolReplaced by GPT-6.1 Sol | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $11.33 |
| GPT-6 Astra | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $27.91 |
| Muse Glimmer 30B | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $3.10 |
| Grok 4.3 | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $4.10 |
| Qwen3.8 27B | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $7.30 |
| gpt-oss-120b | Pass | Pass | Pass | Pass | Pass | Pass | 6 of 6 | $0.422 |
| DeepSeek V4 Pro | Pass | Fail | Pass | Pass | Pass | Pass | 5 of 6 | $5.14 |
| Gemini 3.8 Flash | Pass | Pass | Fail | Pass | Pass | Pass | 5 of 6 | $7.17 |
| Qwen3.8 Max | Pass | Pass | Fail | Pass | Pass | Pass | 5 of 6 | $16.32 |
| Claude Opus 5Replaced by Claude Opus 5.5 | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $16.89 |
| GPT-5.5Replaced by GPT-6 Astra | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $17.02 |
| Claude Fable 5.1Replaced by Claude Fable 5.2 | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $33.11 |
| GPT-5.6 TerraReplaced by GPT-6.1 Terra | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $6.68 |
| MiniMax M3 | Pass | Pass | Pass | Fail | Pass | Pass | 5 of 6 | $1.82 |
| Claude Sonnet 4.6Replaced by Claude Sonnet 5.5 | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $11.24 |
| Kimi K2.6Replaced by Kimi K3.1 Pro | Pass | Pass | Pass | Pass | Pass | Fail | 5 of 6 | $11.69 |
| GPT-5.6 LunaReplaced by GPT-6 Luna | Pass | Fail | Pass | Pass | Pass | Fail | 4 of 6 | $0.677 |
| Gemini 3.5 Flash-Lite | Pass | Fail | Pass | Pass | Pass | Fail | 4 of 6 | $2.47 |
| Gemma 4 31B | Fail | Fail | Pass | Pass | Pass | Fail | 3 of 6 | $0.558 |
| Gemini 3.1 Flash-LiteReplaced by Gemini 3.5 Flash-Lite | Fail | Fail | Pass | Pass | Pass | Fail | 3 of 6 | $1.41 |
| GPT-5.4 miniReplaced by GPT-6.1 Terra | Fail | Fail | Pass | Pass | Pass | Fail | 3 of 6 | $3.35 |
| Claude Haiku 4.5Replaced by Claude Haiku 5.5 | Fail | Fail | Pass | Pass | Pass | Fail | 3 of 6 | $6.62 |
| GPT-5.4 nanoReplaced by GPT-6 Luna | Fail | Fail | Pass | Pass | Pass | Fail | 3 of 6 | $1.51 |
| Llama 4 Maverick | Fail | Fail | Pass | Fail | Pass | Fail | 2 of 6 | $0.665 |
| Mistral Medium 3.5Replaced by Mistral Large 4 | Fail | Fail | Pass | Fail | Pass | Fail | 2 of 6 | $7.40 |
| Ministral 3 14B | Fail | Fail | Pass | Fail | Pass | Fail | 2 of 6 | $1.07 |
| Mistral Small 4 | Fail | Fail | Pass | Fail | Pass | Fail | 2 of 6 | $1.33 |
The tasks
Each page shows the task as the models saw it, the traps in it, every check in words and what each model got wrong.
- Categorise an Italian bar's bank statement and total its movements
- Italian. 27 of 36 runs passed.
- Categorise a café's monthly bank statement and total its operating costs
- English. 24 of 36 runs passed.
- Extract a two-page supplier invoice for accounts payable
- English. 34 of 36 runs passed.
- Extract an Italian supplier invoice with an out-of-scope expense line
- Italian. 31 of 36 runs passed.
- Compute two Italian professional invoices with pension contribution, IVA and withholding tax
- Italian. 36 of 36 runs passed.
- Work out a cleaner's weekly gross pay with breaks, overtime and Sunday rates
- English. 19 of 36 runs passed.
Hand this work to an agent
An AIGROW agent does work like this inside your own tools, against rules you approve in writing, and escalates what the rules do not cover instead of guessing.
October 2026 edition