Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

You can cut your Amazon Bedrock bill without switching models: Flex costs half as much as Standard

Your Amazon Bedrock bill can keep climbing even after you trim prompts, because the request settings set the price. For background jobs like overnight reports, the Flex tier on Nova 2 Lite costs half as much as the regular Standard tier. Switching the reasoning mode from high to low cut an agent's response on Claude Opus 5 from 2,521 to 936 tokens. Change these settings one at a time, checking both your bill and output quality.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
You can cut your Amazon Bedrock bill without switching models: Flex costs half as much as Standard
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1Capping the response at 155 tokens instead of 2,048 saved 1,893 tokens
Key metric #2A one-hour cache pays off from the second request per hour when pauses run longer than 5 minutes
Key metric #3Flex is not supported on Claude Sonnet 5 and Opus 5: the request fails with an error

You have already shortened your prompts and cut unnecessary calls, yet your Amazon Bedrock bill for running models keeps growing. The AWS team showed that part of the money goes to four parameters that are set in every request: how long the cache lives (the remembered part of a prompt that does not have to be paid for again at the full price), which service tier is selected, how much the model "thinks" before answering, and what ceiling is placed on response length. You do not need to change the model or the prompt text.

For example, an overnight report does not need chat-level speed, and if you run it on the Flex tier, it costs half as much as the regular Standard tier. And if an agent spends 936 tokens instead of 2,521 on its response, the bill for it falls almost threefold. A token here is the unit of text you pay for.

Cache: a longer lifetime is not always cheaper

If system instructions, tool descriptions, or reference context repeat, you can cache them. For supported Anthropic models, cache entries can live for 5 minutes or 1 hour. For Claude Sonnet 5 and Opus 5, writing a 5-minute entry costs 1.25 times the normal input price, writing a 1-hour entry costs 2 times it, and reading from the cache costs just 0.1.

The arithmetic is simple. If requests with the same beginning arrive more often than once every 5 minutes, the short cache pays for itself from the first repeat. If the pauses are longer, it gets rewritten on every request at 1.25 times the price. In that case, the one-hour cache is cheaper starting from the second request within an hour. To make the "no cache" option break even, you would need three requests per hour.

There are also pitfalls. Any change to the beginning of a prompt, whether to a system instruction or a tool description, causes a miss and a new write. Validation runs in the order "tools, system, messages," so an edit near the start invalidates everything that follows it. In a test on Sonnet 5, two such edits across six calls led to zero cache reads and a new write. The cache is tied to the account and region, not to your individual clients, so in applications serving many clients, teams add a SHA-256 hash of the client to the start of the prompt. If you need strict data isolation, keep clients in separate AWS accounts.

Flex tier and reasoning effort

Background jobs, evaluation runs, and overnight summarization do not need the speed of an interactive chat. For these, some models offer the Flex tier. AWS tested it on Nova 2 Lite: 200 requests on Standard and 200 on Flex, sent in pairs. The token counts were equal, all 400 requests received the requested tier, and the difference in median latency was indistinguishable from noise. The published Flex price in the us-west-2 region is half the Standard price, and the bill confirmed it.

Support depends on the model version: Nova 2 Lite and gpt-oss-120b have Flex, while Claude Sonnet 5 and Opus 5 do not. A request to an unsupported model fails with a ValidationException and does not silently fall back to Standard. If your policy allows a fallback, catch exactly that error and retry once with the default tier. Do not catch every ValidationException, because malformed requests also raise it.

Reasoning is more expensive. "Thinking" tokens are billed as output tokens. Claude Sonnet 5 and Opus 5 run adaptive thinking at a high level by default, even when the thinking field is not specified. An application migrated from Sonnet 4.6 without changes will start paying for thinking, and maxTokens now limits thinking and response text together.

AWS ran an agent with tool calls on Claude Opus 5: three tasks, eight runs at each effort level. All 24 runs at every level were correct. The median response size grew from 936 tokens on low to 1,484 on medium (+59%), 2,521 on high (+169%), 2,964 on xhigh, and 3,762 on max (+302%). On code generation, the difference between low and high (676 versus 743 tokens) fell within the spread between runs. The caveat is an honest one: near the edge of a model's capabilities, quality at low effort may drop, so you should measure tokens together with the share of tasks completed.

Response ceiling: insurance, not protection

maxTokens limits the response that a bad prompt, a regression, or a loop could inflate. When a request starts, Bedrock reserves input tokens plus maxTokens against the quota, so an overly high limit reduces the number of concurrent requests. In a test on Sonnet 5, normal responses reached 103 tokens, and the limit was set to 155. A deliberately wordy request with a 2,048 ceiling produced 2,048 tokens, while with 155 it produced only 155, saving 1,893 tokens. Both responses were truncated, so the limit saves money but does not make a bad prompt useful. In multi-step chains, a truncated response can break everything downstream.

In practice, teams set the limit based on the P95 or P99 (the length that 95% or 99% of responses fit within) for each task type and add headroom to account for thinking. Stops due to max_tokens should be monitored, and the request should be investigated before the limit is raised. The limit is not a defense against harmful content: for that, there is Bedrock Guardrails.

What to do

The procedure is the same for all four parameters: check model support, change one setting, measure tokens or cost, then check quality. If you route requests to different models, keep a capabilities card for each model version, region, and API covering cache lifetimes, tiers, reasoning levels, and limits. Bedrock's response returns the actual tier and cache counters, and these should be reconciled against CloudWatch and the bill.

Story

Key Questions Answered

What are the key details on Cache: a longer lifetime is not always cheaper?

If system instructions, tool descriptions, or reference context repeat, you can cache them. For supported Anthropic models, cache entries can live for 5 minutes or 1 hour. For Claude Sonnet 5 and Opus 5, writing a 5-minute entry costs 1.25 times the normal input price, writing a 1-hour entry costs 2 times it, and reading from the cache costs just 0.1. The arithmetic is simple. If requests with the same beginning arrive more often than once every 5 minutes, the short cache pays for itself from the first repeat.

What are the key details on Flex tier and reasoning effort?

Background jobs, evaluation runs, and overnight summarization do not need the speed of an interactive chat. For these, some models offer the Flex tier. AWS tested it on Nova 2 Lite: 200 requests on Standard and 200 on Flex, sent in pairs. The token counts were equal, all 400 requests received the requested tier, and the difference in median latency was indistinguishable from noise. The published Flex price in the us-west-2 region is half the Standard price, and the bill confirmed it.

What are the key details on Response ceiling: insurance, not protection?

maxTokens limits the response that a bad prompt, a regression, or a loop could inflate. When a request starts, Bedrock reserves input tokens plus maxTokens against the quota, so an overly high limit reduces the number of concurrent requests. In a test on Sonnet 5, normal responses reached 103 tokens, and the limit was set to 155. A deliberately wordy request with a 2,048 ceiling produced 2,048 tokens, while with 155 it produced only 155, saving 1,893 tokens.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…