Amazon Bedrock adds server-side web search, Microsoft assigns token budget targets to divisions, and Google opens Gemini in Classroom to students of all ages—three steps that treat generative AI as operated infrastructure.
Artificial Intelligence··Morning
Server-side web search inside Bedrock
AWS announced general availability of Web Search on Amazon Bedrock. The tool grounds responses in current web knowledge without a third-party provider and is enabled with one tools entry on the OpenAI-compatible Responses API. Per the AWS Machine Learning Blog, search runs server-side: the model flags a need for current knowledge, Bedrock forms the query, pulls content from an Amazon-operated index of billions of refreshed documents and a built-in knowledge graph, and injects snippets with source URLs and titles into context. Responses return structured citations. AWS says it selects query-relevant passages rather than raw pages, and data by default stays in the environment. Authorisation uses a short-lived SigV4 bearer token from existing AWS credentials, valid up to 12 hours. Callers need at least bedrock-websearch:InvokeSearch; InvokeFetch allows full-page reads and ExternalWebAccess enables live-web retrieval. If InvokeSearch is denied, Web Search is off and the model answers from training data. Only indexed-web retrieval is live today; live-web retrieval comes later. At launch the tool supports OpenAI models on Bedrock.[1]
A token target for each Microsoft division
On the same day, 404 Media reported an email from Microsoft executive vice president Jay Parikh. He told employees to watch token use and said maximising tokens is not the company's goal. Microsoft made OpenAI's GPT-5.6, cheaper than other models, the default internal model to get more value from token spend. Linked Copilot guidelines say that as of July 2026 each division will have an AI token budget target and staff can track their own spending. No target figure is shared, but many engineers spend from hundreds of dollars a month to a few thousand dollars in tokens, and further limits may follow. Parikh writes that the firm optimises impact per token, not token volume, and does not want to slow the AI-first push. 404 Media says the shift did not stem from a cash crunch: latest earnings showed revenue, operating income and net income rising year on year. The site has also reported Amazon, Adobe, Atlassian and Citi trying to throttle employee AI use. An unnamed Microsoft employee found it striking that a firm hosting AI infrastructure is telling staff to curb spending. Microsoft did not immediately comment.[2]
Gemini for every classroom age, and the shared direction
Google Workspace Updates said the Gemini tab in Google Classroom will open to K-12 and higher education students of all ages from 10 August 2026. Rollout starts on the web on 10 August and on mobile on 17 August, taking one to three days to appear, covering Education Fundamentals, Standard and Plus. Google launched Gemini in Classroom for educators last year and later opened it to higher education students aged 18 and older. Students can build flashcards or practice quizzes from course materials; Gemini Notebook users can also create study guides, audio overviews and infographics. Starter prompts can pull class titles, assignment instructions and curriculum materials from Classroom. The Gemini tab is on by default for teachers and students of all ages under the Gemini in Classroom setting; if already off for students it stays off for all, including those under 18. Google advises admins who want to block access for users under 18 to place those users in a separate organisational unit and turn access off there. Together the three moves show generative AI shifting from trial features into operated services: search permissions in Bedrock, division budgets at Microsoft, and admin age-scoped access in Classroom. Shared is that design now covers capability and who may use which resource under which budget and management setting.[3], [1], [2]
Related columns
For more information on this topic, you can read the related columns.
