mission-finops --bedrock-cost-model

Bedrock total cost model.

Token pricing is one line on a Bedrock bill. The surrounding costs, the Knowledge Base capacity floor, invocation logging, committed throughput, agent call fan-out, are where the bill lands well above the token estimate. This model puts all of them on one screen and shows how much of the total is not tokens.

Every rate below is an editable input with a starting default. Enter the prices from the AWS pricing pages for your models and region, then read the itemised total.

Rates entered below are defaults last verified on August 30, 2026. They are starting points, not authoritative prices, and several are region-specific. Confirm each against the Amazon Bedrock pricing page, the OpenSearch Serverless pricing page, and the CloudWatch pricing page for your region before relying on any figure. Results are estimates based on the values you enter.
How to read this. The headline is total monthly cost, not token cost. The line directly under it is the ratio of non-token cost to token cost: when that ratio is high, a token-only estimate would have badly understated the bill, and the decision the output supports is whether the surrounding infrastructure, not the model choice, is what needs attention. The breakdown is sorted largest line first so the thing to look at is at the top.
Inference baseline

The token line. Enter your own per-million prices from the Bedrock pricing page for the model and region you use. There is no built-in model price list.

Average prompt size, including any retrieved context passed in.
Average completion size.
The number of end-user requests, before any agent fan-out below. Lower-volume workloads carry the same fixed capacity floor, which is where the surrounding costs dominate.
From the Bedrock pricing page for your model. Region-specific.
From the Bedrock pricing page for your model. Region-specific.
Agent call multiplication

Agent and tool-use workflows turn one user request into many model calls. This is the most common source of underestimation: an agentic pattern commonly consumes several times the tokens a per-request estimate assumes.

1 for a single call. Raise it for retrieval loops, tool calls, reflection, and multi-step agents. Multiplies the token cost only.
Knowledge Base and RAG infrastructure

Amazon OpenSearch Serverless charges a minimum provisioned capacity floor per OCU-hour whether or not the Knowledge Base is queried. This standing cost is independent of token volume and is the most commonly missed Bedrock line. The minimum OCU count has changed over time, so it is a field here, not a constant. Verify the current minimum.

The number of OpenSearch Compute Units held. Enter the current minimum, or more if you have scaled up. Not hardcoded.

Confirm the current minimum at OpenSearch Serverless capacity limits.

The per-OCU-hour price for your region. Charged 730 hours a month regardless of query volume.
Storage held for the vector index.
Per-GB monthly storage price for your region.
Tokens sent to the embedding model when ingesting or re-indexing source content.
From the Bedrock pricing page for your embedding model.
Observability: model invocation logging

Invocation logging writes full prompts and completions to CloudWatch Logs or S3. On a high-volume workload, verbose logging can rival inference cost. Volume is derived from average prompt plus completion size and request count.

Prompt plus completion plus metadata, as written. A rough guide is a few bytes per token of combined prompt and completion.
CloudWatch Logs ingestion price for your region.
Archived-log storage price for your region.
How long logs are kept. Storage cost scales with retention.
Committed capacity and overflow

Bedrock committed capacity is reserved throughput in tokens per minute for a fixed term. Traffic above the reservation spills to on-demand rates rather than being throttled. Enter the reservation cost and the covered share, and the model shows the utilization point below which the reservation loses money. The default reservation cost is sized to the sample request volume above, so toggling this on shows a sensible reservation rather than an oversized one. Raise it well past the covered token value to see what over-committing looks like.

Discounts and reductions

Batch inference and prompt caching reduce token cost on eligible volume. Enter the share of volume eligible and the reduction each applies. The batch share and the cache share are applied independently and compound if they overlap, so set each to reflect non-overlapping volume.

Batch inference discounts eligible token volume. Set the share of your volume that runs as batch.
The percentage reduction batch inference applies to eligible volume. Confirm on the Bedrock pricing page.
Cached input tokens are billed at a reduced rate. Set the share of input tokens that hit the cache.
The percentage reduction applied to cached input tokens. Confirm on the Bedrock pricing page.
Data transfer and fine-tuning

Data transfer out and, if you customize a model, training, custom model storage, and inference on the customized model.

Outbound data leaving AWS for this workload.
Region-specific outbound transfer price.
One-time or periodic training run cost. Enter 0 if not fine-tuning.
Monthly storage for a customized model.
Inference on the customized model, if priced separately from base token rates.

Total monthly cost

--

--

These are estimates based on the values you enter, not quotes or authoritative prices. Several inputs are region-specific and the AWS pricing structure changes. Confirm every rate on the relevant AWS pricing page before relying on any figure.

Method and sources

Token cost is input tokens plus output tokens per request, multiplied by requests per month and by the agent call multiplier, at your entered per-million prices, then reduced by any batch and prompt-cache shares. The Knowledge Base floor is provisioned OCUs times the OCU rate times 730 hours, plus vector storage and embedding ingestion. Invocation logging derives gigabytes from bytes per call and call volume, priced at ingestion plus storage across the retention period. Committed capacity is the reserved monthly cost plus on-demand pricing on the overflow share. The total is the sum, sorted largest first, and the non-token to token ratio is reported beneath it.

Last reviewed: August 30, 2026.

Independent educational guidance from Mission FinOps. Not affiliated with or endorsed by Amazon Web Services. This is not financial advice.