Bedrock Model Fine Tuning Costs

When fine-tuning a model in AWS Bedrock, you must purchase provisioned throughput to use your customized model for inference, regardless of which base model you choose.

For your specific scenario with 3 epochs and 1 million training tokens, your costs would include:

1. Fine-tuning training cost: Based on the number of tokens in your training data multiplied by the number of epochs (1M tokens × 3 epochs)
2. Custom model storage cost: Monthly fee for storing your fine-tuned model
3. Provisioned throughput cost: Required for inference with your custom model

Provisioned throughput options include:

• No commitment option: Can be turned off when not in use and is billed by the hour
• Longer-term commitments: 1-month or longer options available at different rates

The pricing varies by model. For example, for Cohere Command, provisioned throughput with a 1-month commitment costs $39.60 per model unit. Without a commitment, you would pay the hourly rate (which would be higher per hour than the committed rate).

For your 8-hour testing period, you could choose the "No commitment" option, which allows you to turn off the provisioned throughput when not in use, minimizing costs during your testing phase.

All fine-tuned models, including those based on Nova Lite or any other base model, require provisioned throughput for inference - there is no "free" option for provisioned throughput with any model.
Sources
Unlocking the Potential of AWS Bedrock: Understanding Customization, Throughput, and Pricing | AWS re:Post
Few-shot prompt engineering and fine-tuning for LLMs in Amazon Bedrock | AWS Machine Learning Blog
Pricing
Security best practices to consider while fine-tuning models in Amazon Bedrock | AWS Machine Learning Blog
Fine-tune LLMs with synthetic data for context-based Q&A using Amazon Bedrock | AWS Machine Learning Blog