Bedrock models with fine-tuning / customization
These Amazon Bedrock models can be customised on your own data — fine-tuning or continued pre-training — producing a private model only your account can invoke.
8 of the models on Bedrock qualify today, from $0.035 per 1M input tokens. Rebuilt daily from the AWS catalogue, so this list does not go stale.
What teams get wrong about this
Serving the result is where the economics change. Most customised models can only run on Provisioned Throughput, billed by the hour whether or not you send traffic, so the floor is a fixed monthly cost rather than a per-token one. The exception is a short list — Amazon Nova Lite, Nova 2 Lite, Micro and Pro in us-east-1, Llama 3.3 70B in us-west-2 — which can be deployed for on-demand, per-token inference if customised on or after 2025-07-16. Outside that list, fine-tuning is rarely cheaper than a good prompt plus retrieval unless you have sustained, high volume.
All 8 models, cheapest first
| Model | Input /1M |
|---|---|
| Nova Micro amazon.nova-micro-v1:0 | $0.035 |
| Nova Lite amazon.nova-lite-v1:0 | $0.060 |
| Llama 3.1 8B Instruct meta.llama3-1-8b-instruct-v1:0 | $0.220 |
| Nova 2 Lite amazon.nova-2-lite-v1:0 | $0.300 |
| Llama 3.1 70B Instruct meta.llama3-1-70b-instruct-v1:0 | $0.720 |
| Llama 3.3 70B Instruct meta.llama3-3-70b-instruct-v1:0 | $0.720 |
| Nova Pro amazon.nova-pro-v1:0 | $0.800 |
| Titan Multimodal Embeddings G1 amazon.titan-embed-image-v1 | — |
Capability flags come from the AWS Bedrock model catalogue; prices from the Price List API. Region counts include cross-region inference profiles. Methodology.
Questions
Does fine-tuning on Bedrock require Provisioned Throughput?
Usually. A custom model built on Amazon Nova (Lite, 2 Lite, Micro or Pro, in us-east-1) or Llama 3.3 70B (in us-west-2) can be deployed for on-demand inference and billed per token, provided it was customised on or after 2025-07-16. Every other custom model needs Provisioned Throughput: model units by the hour or with a term commitment. Budget for that hourly floor before the token cost.
Is fine-tuning worth it versus prompting?
Usually only at sustained volume, or when you need behaviour that prompting cannot reach — a house style, a rigid output format, or a domain vocabulary. For factual grounding, retrieval beats fine-tuning: it is cheaper, and you can update it without retraining.