AWS Bedrock cost: per token on-demand, or provisioned throughput

Quick answer

Bedrock bills on-demand per 1,000 input tokens and per 1,000 output tokens, with rates that vary by model (larger, more capable models cost more), plus a provisioned-throughput option billed per model-unit-hour for steady, high-volume use. Prompt length, output length, and model choice drive on-demand cost, so concise prompts, capped outputs, and the smallest model that works are the levers.

Amazon Bedrock provides access to foundation models through one API. Its pricing follows the token model common to generative AI: you pay for the tokens going in (your prompt and context) and the tokens coming out (the model's response), at rates that depend on which model you call.

On-demand: per token, by model

On-demand, you pay per 1,000 input tokens and per 1,000 output tokens. Rates vary widely by model: larger, more capable models cost more per token than smaller, faster ones. Output tokens are often priced higher than input tokens. There is no minimum, so on-demand suits variable or exploratory usage.

Provisioned throughput for steady volume

Mode	Bills for	Best when
On-demand	Per 1,000 input + output tokens	Variable, exploratory, low volume
Provisioned throughput	Per model-unit-hour	Steady, high, predictable volume
Batch	Discounted per-token for async jobs	Large offline workloads

Provisioned throughput reserves model capacity for a term, billed per model-unit-hour, which is cheaper per token at steady high volume and guarantees throughput. Batch inference offers a discount for large asynchronous jobs that do not need real-time responses.

The token levers

On-demand cost is token count times rate, so the controls are: keep prompts and context concise (long system prompts and stuffed context inflate input tokens on every call), cap output length to what you need, cache or reuse responses where inputs repeat, and choose the smallest model that meets the quality bar rather than the largest by default. Retrieval that sends only relevant context, rather than everything, cuts input tokens sharply.

Controlling Bedrock cost

Right-size the model to the task, trim prompts and context, cap outputs, use batch for offline work, and move to provisioned throughput only once volume is steady and high enough to beat on-demand. The token discipline is the same across every generative-AI service, including Azure OpenAI and Vertex AI model APIs.

FAQ

How is AWS Bedrock priced?

On-demand per 1,000 input tokens and per 1,000 output tokens, with rates varying by model (larger, more capable models cost more, and output is often priced above input), plus a provisioned-throughput option billed per model-unit-hour for steady high volume, and discounted batch inference for asynchronous jobs.

How do I reduce Bedrock cost?

Keep prompts and context concise since long inputs inflate cost on every call, cap output length, cache or reuse responses where inputs repeat, choose the smallest model that meets the quality bar, and send only relevant retrieved context. For steady high volume, provisioned throughput beats on-demand.

What is Bedrock provisioned throughput?

A mode that reserves model capacity for a term, billed per model-unit-hour rather than per token, which is cheaper per token at steady, high, predictable volume and guarantees throughput. It suits production workloads with consistent demand; on-demand suits variable or exploratory use.

Does model choice affect Bedrock cost?

Significantly. Larger, more capable models cost more per token than smaller, faster ones, sometimes by an order of magnitude. Choosing the smallest model that meets your quality requirement, rather than the largest by default, is one of the biggest Bedrock cost levers.

Is batch inference cheaper on Bedrock?

Yes, for large asynchronous workloads that do not need real-time responses, batch inference offers a discount over on-demand per-token pricing. Use it for offline processing of big datasets where latency does not matter, and reserve on-demand for interactive use.

Does C3X estimate Bedrock cost?

Bedrock cost is driven by token volume and model choice, which are usage inputs. C3X prices the surrounding infrastructure, and you model prompt and output token volumes per model to estimate the generative-AI charges.

What to do next

Estimate the infrastructure around your AI workload before you build it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.