AWS tests show how reasoning effort and service tiers change Bedrock costs

Image: Primary

Published Thursday, November 12, 2026 · 4:09 PM EST

AWS reported that Claude Opus 5 used 169% more median output tokens at high reasoning effort than at low effort in a Bedrock agent test, with all 24 runs correct at each level. Reasoning tokens are billed as output tokens, making the setting a direct cost variable. In a separate Nova 2 Lite test, AWS said Flex requests cost half the Standard rate, with equal token counts. The findings depend on workload and model support: the code-generation test showed no reliable token difference between low and high effort, and unsupported Flex requests fail rather than fall back automatically.

Sources

AWS Machine Learning Blog

In this story

Anthropic Amazon

Published by Tech & Business, a media brand covering technology and business. This story was sourced from AWS Machine Learning Blog and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

Products

Microsoft 365 outage disrupts email, collaboration and AI connectors

Microsoft 365 connectivity problems disrupted multiple services Thursday, with Microsoft identifying elevated infrastructure load across multiple regions, CNET reported. Its service alert named Exchange Online, Teams, SharePoint Online, Copilot 365 and Microsoft Defender XDR.

Security

Researchers report 21 N-central servers compromised through new exploit

Attackers have compromised at least 21 production N-able N-central servers through a newly disclosed exploit chain, security company Huntback told BleepingComputer. Nineteen ran the latest Hotfix 2 release, issued November 10. The attacks installed XMRig cryptocurrency miners.

Capital AI

Orakl Oncology raises $34M to expand cancer drug prediction platform

Paris-based Orakl Oncology has raised a $34M Series A co-led by Norrsken VC and Singular, bringing total funding to $50M.

Capital Infrastructure

Mount Thor raises $45 million for Apple silicon cloud infrastructure

Mount Thor announced $45 million in Seed and Series A funding to expand cloud infrastructure built on Apple silicon. CRV and Conviction co-led the Series A; Chad Byers and Offline Ventures co-led the Seed.

Capital

Nebuly raises $21 million to measure the economic value of AI agents

Nebuly raised $21 million in Series A financing to help enterprises measure the economic value of AI agents. Planven led the round for the startup, which is based in Turin and New York.

Capital Infrastructure

Gama raises €15M to move satellite solar arrays into volume production

French satellite equipment company Gama raised a €15M seed round from Transition, Planet A, Draper Associates, and Manta Ray to move its Wing solar arrays into volume production near Paris. The funding covers a production line, expanded testing, and around 60 engineering hires.