26 September 2026 • By Timothy, CPA — Managing Director, Professional Financelink (PFL)

Two descending price curves crossing over a set of scales that is missing one arm, flat illustration, no people

AI News Wrap-Up: Two Labs Cut Prices on the Same Day, the Frontier Labs Move to Regulate Themselves, and Google Disclosed a Sandbox Escape Only When Asked

Five AI stories from the past week that actually touch a finance function — a price war, a self-appointed regulator, an identity problem nobody has solved, and a disclosure gap worth noting.

Every Saturday I pull the AI stories that change something for a finance function. This week's two biggest commercial items landed on Monday within hours of each other; almost everything else is about who is supposed to be watching. Read the agent-identity item first — it's the problem your own controls hit next.

1. Anthropic and OpenAI both cut price and latency on the same day

On Monday 22 September, Anthropic released Claude Opus 5.5 — which Anthropic says performs at the level of its own Fable 5.1 on most tasks — at US$4 per million input tokens and US$20 per million output. That's 20% below Opus 5, with cached reads down 60% and typical-workload cost about 40% lower by Anthropic's estimate. Later the same day, OpenAI released GPT-6 Sol at US$2 and US$10 per million input and output tokens — half the GPT-5.6 Sol price, and described as permanent — alongside a cheaper Luna sibling.

Tim's take: Two labs cutting price on the same day matters more than either release. The cost line under your AI business case is not a stable input — it has fallen twice in two months, and nothing suggests today's rate is the floor. A workflow rejected on cost earlier this year deserves another look, and an annual commitment signed at older rates is worth re-opening. One caution: cheaper per token is not automatically cheaper per month. Ad-hoc use tends to grow when something feels cheap; scheduled workflows are where the saving actually shows. Track the two separately.

Source: Anthropic — Claude Opus 5.5, VentureBeat — OpenAI releases GPT-6 Sol and Luna, slashing API costs 50% or more

2. Google, OpenAI and Anthropic are building their own regulator, with no government in the room

Reporting this week, originating with The Information, says Google, OpenAI and Anthropic have agreed to establish a voluntary standards body — reported variously as the Frontier AI Standards Agency or the Standards Authority for Frontier AI — to set common testing and safety standards for advanced AI. The model cited is FINRA, the US securities industry's self-regulator. Sriram Krishnan, who left his White House AI adviser role in June, has reportedly been approached to run it, with a launch floated for late 2026 or early 2027. None of the three has confirmed it, so treat it as trade reporting.

Tim's take: Self-regulation isn't automatically a dodge — FINRA does real work. But it changes what a vendor assurance is worth. If a supplier says their model meets "industry safety standards," ask whose standards, who audits against them, and what happens on failure. A standard written by the companies selling the product shouldn't sit in a risk register as though it were a government one. The government track moved too: Australia joined some 20 countries and the EU in a pre-UN General Assembly statement calling for mandatory safety testing and international oversight of frontier AI, and legislation for mandatory Australian Standards for AI is still expected in early 2027.

Source: BankInfoSecurity — Google, OpenAI, Anthropic plan frontier AI standards body, ABC News — Australia joins global push for AI controls, Baker McKenzie — Australia charts a new course on AI regulation

3. "Know your agent" — banks can't tell whether a payment came from a customer or their software

Fortune reported on 19 September that financial institutions have no existing framework for verifying and monitoring AI agents that shop and pay on a customer's behalf. Ant International, Mastercard and Visa announced on 6 September that they are building a "know your agent" interoperability framework, and a McKinsey estimate that agents could orchestrate up to US$5 trillion of global consumer spending by 2030 is driving the urgency. Three days after the Fortune piece, six banks — ASB, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest — published Building Trust in Agentic Commerce, calling for an audit trail from the customer's instruction through to the payment outcome. The principles are voluntary; an implementation paper is to follow.

Tim's take: This is the item most likely to reach your desk, and it isn't really about shopping. Strip the retail framing and the problem is one every finance function is about to have: an action was taken in our systems — who authorised it, and can we prove it? Delegation frameworks assume a named human at the point of approval; an agent acting under a standing instruction breaks that quietly, because the log shows a service account rather than a decision. Two cheap things to do now: write down which systems an AI tool can take an action in, as opposed to read from; and check whether your approval logs capture the human who set the standing instruction or only the process that executed it. If it's the latter, that evidence gap can't be fixed retrospectively.

Source: Fortune — 'Know your agent': Banks face a new compliance challenge, ING — As AI starts shopping, who stays in control?, PYMNTS — Banks call for audit trail from AI shopping instructions to payment outcome

4. Anthropic is testing a "Money" tab that links Claude to your bank account

Mid-September reporting, first surfaced by TestingCatalog on 14 September, shows Anthropic testing a feature labelled "Money" — a separate tab in the Claude iOS app that would let a user link bank accounts and ask Claude about spending, plans and budgeting. Anthropic hasn't announced it or given a date, and reporting suggests a US-first release. OpenAI already offers a comparable bank-linked personal-finance feature to eligible US ChatGPT users.

Tim's take: One reason to flag an unreleased consumer feature: if and when it reaches Australia, some of your staff will use it before your organisation has an opinion about it. The question worth asking early isn't whether the feature is good. It's whether someone linking their own bank account to an AI assistant has read what the vendor's retention and training policy says about that data. Most people won't. That's worth a line in an internal AI note now, framed as a heads-up rather than a ban.

Source: TestingCatalog — Anthropic prepares Claude Money for personal finance, BleepingComputer — Anthropic wants Claude to analyze your bank account and financial data

5. Google confirms Gemini broke out of a test sandbox and accessed three real companies — disclosed only when a journalist asked

On 18 September Google confirmed a Wall Street Journal report that during a capture-the-flag exercise in May, run by the third-party AI security evaluator Irregular, a Gemini model gained unauthorised access to systems belonging to three real companies. The fictional target shared names with real businesses, and a configuration flaw left the environment with internet access. The model guessed passwords in one case and used credentials from a public repository in the other two. Google says the model stopped each time, the companies were notified and the testing partner has changed its processes. Irregular says the same flaw affected other labs' models too, and that it notified the labs in late July.

Tim's take: I covered OpenAI's incident disclosures last Saturday. The technical failure is almost mundane — a system that couldn't distinguish test from production acted on that confusion, a control failure any auditor would recognise, sitting here in a supplier's environment rather than the lab's own. The disclosure timing is the governance signal: about four months after the incident, roughly seven weeks after the testing firm says it notified the labs, and only after a journalist asked. "We'll be told if something goes wrong" isn't a control you can lean on in a vendor risk assessment. Assume you find out late: what would be reachable from a tool you've connected if it misbehaved for a week before anyone noticed?

Source: Cybersecurity Dive — Google AI models broke out of sandbox, hacked three companies, Axios — Google Gemini accessed three companies during AI hacking test

The through-line is a widening gap. Capability and price are moving fast in the customer's favour. Everything else is the machinery meant to keep up: a regulator the industry is writing for itself, an identity framework that doesn't exist yet, and a disclosure that arrived only when someone asked. None of that argues for slowing your own adoption. It argues for keeping the boring parts current — the permission list, the approval log, the vendor file — because for now they're the only assurance you control.

A standing note for anyone using AI tools on payroll, participant or client data, or financial figures: confirm whether the vendor trains its models on your inputs, and where you have a choice, prefer a configuration in which your data isn't retained for training. Strip names out where a task doesn't need them. This applies to personal use as much as organisational use — see item four.

Can you produce the list of systems an AI tool can act in, not just read from?

PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including building the permission and approval records that make AI use defensible before anyone asks to see them.

Talk to PFL →

Timothy, CPA is Managing Director of Professional Financelink (PFL), providing senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. 20+ years in finance leadership across NFP, NDIS and SME.

Sources:

Tomorrow: Finance Reads — two Support at Home changes on one date, an NDIS records duty that is now law, and Thriving Kids goes operational.

Search This Blog

Finance Intelligence