Mistral AI’s La Plateforme for Accounting
Automation: Features and Pricing
May 2, 2025
1 Model Selection
The following models are recommended for your use case:
• Invoice OCR:
– Model: Mistral OCR (mistral-ocr-latest)
– Description: A state-of-the-art OCR API that processes images and PDFs,
extracting structured text, tables, and images. It supports multilingual doc-
uments and complex layouts, with a throughput of up to 2,000 pages per
minute.
– Why Suitable: Ideal for extracting structured data from invoices, enabling
automation of fields like invoice number, date, and total amount. Its high
throughput supports your 65,000 invoices monthly.
• Expense Categorization and Fine-Tuning:
– Model: Ministral 8B
– Description: An 8-billion-parameter LLM optimized for edge deployment
and low-latency tasks like text classification. It offers high efficiency for
resource-constrained environments and supports a 32k token context window.
– Why Suitable: Efficiently classifies transaction descriptions into categories,
suitable for processing 80,000 transactions monthly with lower costs than
larger models. Fine-tuning capabilities enhance accuracy for your specific
expense categories. (Mistral AI Models, https://mistral.ai/models)
• AI Agents:
– Model: Mistral Large 2 (mistral-large-2407)
– Description: A 123-billion-parameter LLM with advanced reasoning, multi-
lingual support, and function-calling capabilities, ideal for building complex
AI agents.
– Why Suitable: Powers sophisticated agents for tasks like customer support,
fraud detection, and workflow automation, with the ability to integrate with
external systems and handle complex instructions.
1
2 Pricing Structure
Mistral AI’s pricing for the commercial tier is usage-based, with costs calculated per unit
of processing. Below is the pricing structure for the selected models and tasks:
• Mistral OCR Pricing (Invoice Processing):
– Standard: $1 per 1,000 pages
– Batch Processing: $1 per 2,000 pages
– Assumption: Batch processing is used to optimize costs.
• Ministral 8B Pricing (Expense Categorization and Fine-Tuning):
– Input Tokens: $0.15 per 1 million tokens (estimated, pending official pricing)
– Output Tokens: $0.45 per 1 million tokens (estimated, pending official pricing)
– Assumption: Each transaction requires 30 input tokens (20 words at 1.5 tokens
per word) and 5 output tokens (category label). Self-hosting may reduce costs
to hardware expenses.
• Mistral Large 2 Pricing (AI Agents):
– Input Tokens: $2 per 1 million tokens
– Output Tokens: $6 per 1 million tokens
– Assumption: Agent interactions vary by use case, detailed below. (Mistral AI
Docs, https://docs.mistral.ai/)
• Fine-Tuning Pricing:
– Minimum Job Fee: $4 per fine-tuning job
– Storage Fee: $2 per month per fine-tuned model
– Training Cost: $0.2 per million input tokens, $0.6 per million output tokens
3 Cost Estimation for OCR and Categorization
The cost is calculated for processing 80,000 transactions and 65,000 invoices per
month, assuming batch processing.
3.1 Invoice OCR Cost
• Volume: 65,000 invoices
• Pages per Invoice: Average of 1.5 pages per invoice:
65, 000 × 1.5 = 97, 500 pages
• Cost with Batch Processing: $1 per 2,000 pages
97, 500
= 48.75 units
2, 000
48.75 × $1 = $48.75 per month
2
• Cost per Invoice:
$48.75
= $0.00075 per invoice
65, 000
3.2 Expense Categorization Cost
• Volume: 80,000 transactions
• Tokens per Transaction:
– Input: 20 words, 1.5 tokens per word:
20 × 1.5 = 30 input tokens
– Output: 5 tokens for category label
– Total: 30 + 5 = 35 tokens
• Total Tokens:
– Input: 80, 000 × 30 = 2, 400, 000 tokens
– Output: 80, 000 × 5 = 400, 000 tokens
• Cost Calculation (Ministral 8B, API-based):
– Input cost:
2, 400, 000
× $0.15 = $0.36
1, 000, 000
– Output cost:
400, 000
× $0.45 = $0.18
1, 000, 000
– Total:
$0.36 + $0.18 = $0.54 per month
• Cost per Transaction:
$0.54
= $0.00000675 per transaction
80, 000
• Alternative (Self-Hosted): Assuming $150 monthly GPU cost, with 25% appor-
tioned to categorization, costs are approximately $37.50 monthly, less cost-effective
for low volumes unless shared across tasks.
3.3 Total Monthly Cost for OCR and Categorization
3
Table 1: Total Cost Summary for OCR and Categorization
Task Volume Total Cost (USD) Cost per Unit (USD)
Invoice OCR 97,500 pages $48.75 $0.00075 per invoice
Expense Categorization 80,000 transactions $0.54 $0.00000675 per transaction
Total $49.29
