AWS Example: A Model With a Token Price Twice as High Costs 65% Less per Task
The per-token price on the Bedrock price list is not the price of the work. In AWS's example, a model with a twice-as-high per-token price processed a customer email for $0.00145, while the "cheap" one did it for $0.0042, making it 65% more expensive. The "cheap" model split the text into 50% more tokens, tripled the length of its answer, and took three turns instead of one. So compare models by the cost of a task on your own data, not by the price list.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Imagine you're choosing a model for an agent that reads customer emails, extracts data from them, and replies with strict JSON. On the Amazon Bedrock price list, one model costs $0.50 per million input tokens and another costs $1.00. The obvious choice is the first one. But in the second part of its series on Bedrock spending, AWS engineers show with an example that the opposite is true: the model with a twice-as-high per-token price completed the task for $0.00145, while the "cheap" one did it for $0.0042. That's 65% cheaper. The reason is that you are billed not for tokens but for completed work: a processed email, fields extracted from an invoice, a reply to the user.
The first part of the series addressed who spent the money and how to limit costs. The second part covers the two remaining questions: why a specific scenario is expensive, and whether an agent pays off compared with humans.
Four Hidden Sources of Spending
AWS highlights four factors that do not appear on the pricing page.
First, the tokenizer. Each model family has its own vocabulary, so the same text or JSON turns into a different number of tokens. If model A is 15% cheaper but your data takes up 40% more tokens with it, model B ends up being the cheaper option. With non-Latin alphabets the gap is even larger: text can split two to three times more than English text.
Second, verbosity. Output tokens usually cost 3–5 times more than input tokens. If a model ignores the instruction "return only JSON" and adds polite explanations, you pay for empty words.
Third, agent loops. When calling tools, a model makes several turns, and each one receives the entire conversation history and tool results again. Four turns with a growing context cost far more than a single pass.
Fourth, reasoning. "Thinking" tokens are billed as output: 2,000 tokens of thought plus 200 tokens of answer produce a bill for 2,200. Some models hide their reasoning but still charge for it. AWS's advice is to keep reasoning turned off until tests show a quality gap that justifies the price.
How the 65% Came About
In the example, the agent reads a 750-word email, extracts entities, checks the account status via an API, and returns strict JSON. Model A (the "cheap" one) uses 1,500 tokens on the first turn versus 1,000 for model B, replies with 450 tokens versus 150, and takes three turns instead of one: it gets confused by the tool schema and corrects its mistakes. In total, that's 6,150 input tokens versus 1,000. The result: $0.0042 for A and $0.00145 for B. The prices in the example are illustrative, so you need to verify them on your own data.
For real work, you also need the cost of a successful task. If a task is retried until it succeeds, the cost is the cost of an attempt divided by the share of successes. If a failed task is handed off to a human, the escalation cost multiplied by the share of failures is added to the cost of an attempt.
For measurement, AWS recommends Bedrock model evaluation for quality, CountTokens for comparing tokenizers without running generation, and the usage block in the model response. One caveat: when caching is enabled, the inputTokens field counts only uncached tokens. The total input volume is the sum of inputTokens, cacheReadInputTokens, and cacheWriteInputTokens, but these are billed at different rates. Cache hit rates drift over time: tools get updated, and in three months you will be writing to the cache more often, so the settings should be reviewed.
There are four levers for reducing costs. Prompt caching eliminates resending the same context and delivers the biggest gains on agent paths. Routing sends simple requests to a cheaper model from the same family. A tool schema and a maxTokens limit control verbosity more reliably than requests written into the prompt text. Batch processing halves the rate for large volumes of single-step tasks, but it does not support tool calling.
How to Prove an Agent Pays Off
Here AWS insists on discipline. First, describe the human process: what triggers the work, who takes part, where the handoffs between roles happen, and what counts as completion. Then measure the baseline before the agent launches: execution time, queue length, share of reopened cases, and cost per task. A baseline reconstructed from memory must not be used, since it systematically flatters the agent.
Next, record every agent run as an experiment entry: an identifier, group (agent or control), prompt version, number of retries and tool calls, tokens, and outcome. Human minutes spent on review and the identifier of the revenue event are especially important: without the first, the agent looks free, and without the second, the money cannot be tied back later.
From this data, a single Athena SQL query yields three figures: the cost of a successful result, the number of retries per success, and human minutes per success. The cost per call flatters agents that run in loops, but the cost per success does not.
For the final assessment, AWS proposes a comparison with a control group and a scorecard of five dimensions: adoption (green zone: more than 70% of users active), quality (more than 90% of tasks completed without rework), efficiency, cost per success, and business impact. Green means scale up, yellow means rework, and red means shut down. The thresholds are a starting point, not a norm. The authors themselves acknowledge that the scorecard only works if the team actually switches off red agents.
What to Do in Practice
For developers: compare models on your own traffic and count the tokens of every turn in a task, not just a single request. For product managers and FinOps: set up a control group in advance and record people's time. The limitation of this material is that the prices and thresholds are illustrative, and some of the levers depend on Bedrock's features. The underlying logic, measuring the cost of success rather than the cost of a call, applies to other platforms as well.
- Nov 21, 2026AG-UI 1.0 gives AI agents a common language for talking to apps: 31 event types
- Nov 21, 2026OpenAI launches ChatGPT Wallet: agents buy without ever seeing your card number
- Nov 21, 2026US Advertisers Spent $70.7 Million on ChatGPT Ads in October, Up 127% From April
- Nov 21, 2026ONNX Runtime vanishes from GitHub: Microsoft hides repositories after package hack
Key Questions Answered
What are the key details on Four Hidden Sources of Spending?
AWS highlights four factors that do not appear on the pricing page. First, the tokenizer. Each model family has its own vocabulary, so the same text or JSON turns into a different number of tokens. If model A is 15% cheaper but your data takes up 40% more tokens with it, model B ends up being the cheaper option. With non-Latin alphabets the gap is even larger: text can split two to three times more than English text. Second, verbosity. Output tokens usually cost 3–5 times more than input tokens.
How the 65% Came About?
In the example, the agent reads a 750-word email, extracts entities, checks the account status via an API, and returns strict JSON. Model A (the "cheap" one) uses 1,500 tokens on the first turn versus 1,000 for model B, replies with 450 tokens versus 150, and takes three turns instead of one: it gets confused by the tool schema and corrects its mistakes. In total, that's 6,150 input tokens versus 1,000. The result: $0.0042 for A and $0.00145 for B.
How to Prove an Agent Pays Off?
Here AWS insists on discipline. First, describe the human process: what triggers the work, who takes part, where the handoffs between roles happen, and what counts as completion. Then measure the baseline before the agent launches: execution time, queue length, share of reopened cases, and cost per task. A baseline reconstructed from memory must not be used, since it systematically flatters the agent.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.