In the World of AI: Week of 29 September 2026, Google Holds Back Gemini 4 Argon

The biggest shifts this week were a privacy deadline for software that influences decisions about people, a new Google model that most businesses cannot touch yet, and a notice that a popular Claude model will switch off in under two months. Here is the week from 30 September to 6 October 2026, with one thing to do about each story.

1. The privacy regulator publishes guidance on automated decisions

The Office of the Australian Information Commissioner (OAIC) released guidance on 30 September 2026 for a new transparency rule that starts on 10 December 2026, according to a summary by law firm Allens. Organisations covered by the Australian Privacy Principles must describe in their privacy policy the kinds of personal information a program uses to make, or substantially contribute to, decisions that could significantly affect a person. They must also describe the kinds of decisions made solely by the program and the kinds where it only helps. Allens notes it is a transparency measure only. It gives people no new right to contest a decision, and the duty sits with the organisation using the information, although it may need details from its software suppliers.

Why it matters: The wording covers more than chatbots. Any tool that scores, filters or ranks people, such as a recruitment screener or a credit check add-on, could fall inside it. Whether your business is covered is a legal question, so get advice. Our answer on whether the Privacy Act applies to AI tools is a starting point.

The deadline is about two months out, which is enough time to ask questions but not to leave them.

This week: List every tool that helps decide something about a customer, applicant or employee, and ask each supplier what personal information it uses.

2. Google announces Gemini 4 Argon, but only a few can use it

On 30 September Google announced Gemini 4 Argon, its first Gemini 4 model, pitched at long software projects, legal and financial analysis and cybersecurity. At launch it is available to trusted cyber defenders through Google's Fairwind Program. Google says it will reach paid API customers and Google AI Ultra subscribers next but gives no date, and it says the model will first be tested with early users while safeguards are refined.

Why it matters: Headlines about a new model rarely mean you can use it today. Tools you already pay for will not change overnight, and a rushed switch can break working processes.

This week: Do nothing with Argon yet. Pick one real task, such as summarising a long supplier contract, so you have a fair test ready for when it arrives.

3. Anthropic sets a retirement date for Claude Sonnet 4.5

Anthropic's platform release notes for 30 September 2026 announce the deprecation of Claude Sonnet 4.5, with retirement on 30 November 2026. This mostly affects businesses that have a developer or agency build tools that call Claude directly. Once a model is retired, requests to it stop working.

Why it matters: A custom quote-drafting tool or email sorter may name this model in its settings, and it would stop working on the retirement date unless someone changes it.

This week: Ask whoever built or maintains your AI tools which model each one calls, and get a written plan for any that use Sonnet 4.5. If you want to understand where automation fits, see what AI can automate in a small business.

4. ANZ warns businesses about impersonation fraud

IT Brief Australia reported on 6 October that ANZ is warning businesses about scams and impersonation fraud. ANZ's Cosi De Angelis said human judgement remains one of the strongest defences. The article cites more than 84,700 cybercrime reports nationally in FY25 from the Australian Signals Directorate's annual report, and says ANZ customer scam losses fell 24% between October 2025 and June 2026. ANZ's advice includes confirming changes to banking details by phone rather than email, requiring two people to authorise large or urgent payments, and reviewing cyber controls before Black Friday and the festive season.

Why it matters: ANZ ties the risk to the festive season, when rushed approvals are more likely, and a changed bank account number is the kind of request its checklist targets.

This week: Agree a rule that any change to a supplier's bank details is confirmed by phone, using a number you already held.

What to do this week

• List the tools that help decide things about people, and ask suppliers what personal information they use.
• Choose one real task to test Gemini 4 Argon on when it becomes available.
• Ask your developer or agency which Claude model each of your tools uses.
• Put a phone-call check on every change of supplier bank details.

If your staff already use AI without a clear rule, the AI policy answer covers where to begin.

Sources

• Allens, "Mandatory automated decision-making disclosure requirements are coming: key takeaways from the OAIC guidance", 2 October 2026: allens.com.au
• Google, "Gemini 4 Argon: our next era of frontier intelligence", 30 September 2026: blog.google
• Anthropic, "Claude Platform release notes", entry of 30 September 2026: platform.claude.com
• IT Brief Australia, "ANZ warns businesses of AI scams, impersonation fraud", 6 October 2026: itbrief.com.au