Amazon Web Services lists the new version of its AgentCore Runtime, the serverless service that runs AI agents on Amazon Bedrock, at $0.0169 per gigabyte-hour of memory, against $0.00945 for version 1, according to its pricing page. That is 79% more per gigabyte-hour, by our arithmetic; CPU rises from $0.0895 to $0.1276 per vCPU-hour, or 43%.
Update, 14 November 2026, 18:25 UTC: AWS adds a direct Web Search API to AgentCore. In a blog post dated 11 November, AWS announced the Web Search API on Amazon Bedrock AgentCore, which it says lets “your own code” search the web “with a single API call.” AWS says it returns ranked results from “a web index that Amazon operates” and runs on “the same search infrastructure that powers Alexa+, Kiro, Quick, and other Amazon products.” Until now, AWS says, web search on AgentCore reached agents through AgentCore Gateway. Per the post, a query can return 1 to 25 results, domain filters hold up to 100 domain names each, the search text can be up to 200 characters, and “There’s no resource to create and no API key to manage.” The post does not give a price for the API.
What AWS offers in return is idle time: in Runtime v2, the page says, “Idle memory is reclaimed automatically in Runtime v2 after 120 seconds”. For the memory bill to come out lower, v2 has to bill at least 44% fewer gigabyte-hours than v1 for the same session, by our arithmetic.
Business Pill · Switching Costs Explained
An evergreen Business Engineer explainer on switching costs: why it is hard to leave an AI supplier. It teaches the general idea only and does not discuss AgentCore.
The key insight: As we read the price list, AWS is charging more per unit and less for idle time: v2 memory is 79% dearer per gigabyte-hour, and only agents that wait long enough for memory to be reclaimed come out ahead on the memory bill.
AgentCore Runtime bills per second for the CPU and memory an agent session actually uses. The page says “CPU scales to zero during I/O wait (waiting for LLM responses, tool / API calls, or database queries) and with the new Runtime (v2), memory is reclaimed throughout the session”.
AWS also lists a committed-baseline option for v2, “launching by October 2026”, at $0.0997 per vCPU-hour and $0.0132 per gigabyte-hour. The baseline memory rate is 22% below v2’s consumption price but still 40% above v1’s, by our arithmetic.
Runtime microVMs consumption prices, v2 against v1, from the AWS AgentCore pricing page read on 14 November 2026. Percentages are our arithmetic.
AWS’s Own Example, Recalculated
The pricing page works through a customer-support agent: 1 million sessions a month, each 10 minutes long with 90% I/O wait, using 60 seconds of CPU and memory that steps from 1GB to 2.5GB. AWS puts the total at $0.006703 per session, or $6,703 a month.
Billed for the same CPU-seconds and gigabyte-seconds at v1 rates, that session would cost about $0.004051, by our arithmetic, so the published v2 figure is 1.65 times higher. The comparison holds only if v1 billed the same memory; v2’s case rests on reclaiming memory that v1 would have kept charging for during long waits.
Why Idle Time Decides the Bill
Agents spend much of a session waiting on models, tools and people, which is the pattern Anthropic describes in its guide to always-on agents. Under v2, a session that holds memory through those waits pays the higher rate on all of it; one that sits idle for more than 120 seconds at a time gets that memory reclaimed.
The same question of who pays for overhead runs through Garry Tan’s harness test, which found the cost of a solved task varied up to 2.18 times on one model depending on the software around it.
The Structural Read
As we read AWS’s worked example, the session spends most of its 10 minutes waiting, which is the case v2 is built for; even so, the published v2 cost is 1.65 times the same usage at v1 rates, by our arithmetic.
As we read the committed-baseline rates, AWS is steering steady workloads to commitments: the baseline memory price is 22% below v2 consumption pricing, though still 40% above v1, by our arithmetic.
The framework opens: “AI has crossed from the training era to the inference era.”
As we read it, AgentCore’s v2 prices show that era reaching the hosting layer: the meter now runs on how long agents wait as well as how hard they work. The framework does not discuss AgentCore.
AWS’s published v2 cost for its customer-support example, against the same CPU-seconds and gigabyte-seconds priced at v1 rates (our arithmetic; assumes v1 billed the same memory).
What Is Not Established
The pricing page does not say when v2 launched, whether v1 will be retired or how many customers run on each version. It gives no v1 version of its worked example, so the v1 comparison above is our arithmetic on AWS’s published rates and assumes the same billed usage. Tech Times first reported the price difference; we read AWS’s page directly.
AWS was not contacted for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
AWS’s AgentCore Runtime v2 lists memory at $0.0169 per gigabyte-hour, 79% above v1, and CPU at $0.1276 per vCPU-hour, 43% above, by our arithmetic. Its memory bill comes out lower only for agents whose idle-memory reclaim cuts billed gigabyte-hours by 44% or more.
A note on sourcing. Prices and the worked example come from AWS’s Amazon Bedrock AgentCore pricing page, read on 14 November 2026. The v1 comparison and all percentages are our arithmetic on AWS’s published rates.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
