Generative AI is reshaping industries, but for any business using it, one question looms large: What does this actually cost? Amazon Bedrock provides access to powerful AI, and its pricing all comes down to 'token economics'.
First, What Exactly Is an AI 'Token'?
Before you can understand
the pricing, you have to understand the currency. In the world of large language models (LLMs), that currency is the token. Think of a token as a small piece of text. It's not quite a word and not quite a character; it's the fundamental unit that an AI model processes. A good rule of thumb is that one token equals about three-quarters of an English word. The sentence 'AI is changing business' might be broken down into tokens like 'AI', 'is', 'changing', and 'business'. Everything you send to the model (the input prompt) and everything it sends back (the output generation) is measured, and therefore billed, in tokens. This is why 'token economics' is the critical concept: it's the financial framework for how AI consumption is measured and managed.
The Core Philosophy: Pay for What You Use
Amazon Bedrock’s primary pricing structure is built on a simple utility model: you only pay for what you consume. There are no flat monthly fees or user-seat licenses for basic access. Instead, your bill is a direct reflection of the number of input tokens you send and output tokens you receive. However, 'pay-as-you-go' isn't one-size-fits-all. Bedrock offers different modes of purchasing this AI processing power, chiefly 'On-Demand' and 'Provisioned Throughput', each designed for different business needs and usage patterns. Understanding the trade-offs between these models is key to managing your AI spend effectively.
On-Demand: The Flexible Starting Point
The On-Demand model is exactly what it sounds like. It offers maximum flexibility with no longterm commitments, making it ideal for developers experimenting with new applications or for businesses with unpredictable, spiky workloads. You are charged a specific rate for every million input tokens and a separate, usually higher, rate for every million output tokens. This approach is simple and elastic, allowing you to scale up or down instantly. However, that flexibility can come at a higher pertoken cost compared to other options, and highvolume usage can be subject to rate limits or throttling.
Provisioned Throughput: Buying Power in Bulk
For businesses with steady, highvolume AI workloads, Provisioned Throughput offers a more costeffective and reliable alternative. Instead of paying per token, you commit to a certain amount of processing capacity (called a 'model unit') for a set term, typically one or six months. You pay a fixed hourly rate for this capacity, whether you use all of it or not. The benefit is a guaranteed level of performance with no throttling and, with a longer commitment, a significantly lower effective cost than OnDemand. This model is built for productionlevel applications where latency and predictability are critical.
Not All Tokens Are Created Equal
A crucial layer of Bedrocks economics is that the price per token varies dramatically depending on which foundation model you choose. Bedrock is a marketplace, offering models from Amazon (Titan), Anthropic (Claude), Meta (Llama), and others. A simple classification task might run efficiently on Amazons own costeffective Nova Micro model, which costs mere cents per million tokens. A complex reasoning task requiring a stateoftheart model like Anthropics Claude Opus could cost over 100 times more for the same number of tokens. This makes model selection one of the most significant levers for controlling costs.
Batch and Other CostSaving Measures
Beyond the two main purchasing models, Bedrock offers other ways to optimize spending. For workloads that are not interactive and can tolerate a delay, such as nightly document summaries or data classification, 'Batch' mode offers a straight 50% discount off the OnDemand price. Additionally, some models support features like prompt caching, which can reduce costs for repetitive queries. Customizing models through finetuning also has its own cost structure, involving onetime training fees and monthly storage costs, with the resulting model typically running on Provisioned Throughput. These tools provide further strategic options for managing the total cost of ownership.
