Last Week in AI: 28 September–4 October 2026

A week of changes worth checking against your own work

This edition covers 28 September–4 October 2026. We selected a model release, a released data-tool integration and an announced change to reusable workplace prompts. The useful question for a UK IT team is what each change would mean for an existing task, its permissions and its owner.

The product facts below come from the named publishers. Our suggested checks are editorial advice; we have not run comparative product tests. Confirm your organisation’s available plans and approved configuration before using business information.

28 September: Sonnet 5.5 gives existing Claude users a new comparison

Anthropic released Claude Sonnet 5.5 on 28 September. It positions the model for bounded everyday tasks, bug fixes and document work, with availability including its own platform, AWS, Google Cloud and Azure. That is a release and a vendor positioning claim, not evidence of better results on your organisation’s tasks.

Our suggestion is to rerun a small, authorised example with a known acceptable answer. For a fictional service team, that might be drafting a support note from sanitised incident facts. Ask the same reviewer to assess accuracy, missing caveats and editing effort. Keep the input and review criteria consistent rather than choosing the most impressive one-off response.

For an API integration, check the migration guide before changing the model identifier. Anthropic specifically says integrations using thinking off need to move to its between_tools setting. A model swap is therefore worth treating as a reviewed change. Do not infer a subscription saving or guaranteed cost reduction from vendor benchmark claims.

30 September: Data Agent Kit moves into general availability

Google Cloud announced general availability of Data Agent Kit, its MCP tools and agent skills for connecting coding agents to Google Cloud data services. The GA release adds BigQuery Graph, Bigtable and Managed Service for Apache Spark support. Google says operations use your IAM permissions, including when impersonating a service account.

This is useful to investigate if your team already supports that data estate. Start with a deliberately limited identity and a non-production question, such as inspecting an approved example schema. Have the data owner agree the boundary and require review before changes. Access to more context is not a reason to grant a tool broader authority than the task needs.

Google says the kit has no additional charge, but the cloud services it uses retain their standard pricing. Set an explicit usage boundary and review the query or operation before running it. Our recommendation is to record both whether the answer is useful and whether the access and spending behaviour stayed within the agreed scope.

30 September: Google sets out the transition from Gems to skills

Google’s Workspace announcement describes reusable skills that can be combined in a prompt, and warns that skills do not currently sync between the Gemini app and Workspace. For organisations, this makes the location and owner of a reusable instruction important: a copy in one surface is not automatically the maintained copy in another.

The published rollout plan starts Workspace Rapid Release on 5 October and Scheduled Release on 19 October, with the Gemini app rollout starting 13 October. Workspace availability is limited to the listed eligible editions and add-ons, and users over 18. Do not read the announcement as proof that every employee has the feature today.

For business and enterprise users, the stated removal of Gems is no sooner than 1 March 2027; affected Workspace Studio Ask a Gem flows would stop working at that transition. Our suggested next step is an inventory of Gems and dependent flows, with owners and sample expected outputs. Check the current migration guidance before planning changes, and test a recreated skill rather than assuming identical results.

Choose one reviewable next step

Pick the development that touches a service you already own. Write down the approved test input, the expected result and who will decide whether the change helps. Keep a useful failure in the record as well as a successful result; either can improve the decision.

If none of these changes fits your current work, leave them on the watch list. A dated announcement is a prompt for an informed decision, not an obligation to add another tool or change a working process.