Insights
What we are seeing this month.
MAS · 9 Nov 2026
The MAS AI guidelines ask for one thing first: an inventory. Most firms don’t have one.
On 7 October 2026 the Monetary Authority of Singapore published its Guidelines on Artificial Intelligence Risk Management for financial institutions. The guidelines say: “These Guidelines take effect on 7 October 2027.” (para 1.8). Sections 5 and 6 — the controls across the AI life cycle and third-party AI — are given until 7 October 2028: “and Section 5 and 6 of the Guidelines by 7 October 2028.”
The first thing the guidelines expect is not a policy and not a committee. It is a list. Para 4.5 asks an FI to “establish and maintain an accurate inventory of AI use cases, systems or models”, with “an appropriate frequency to update the inventory”. Para 4.12 then asks the firm to assess each use case by impact, complexity and reliance.
In our sample Diagnostic of a fictional 80-person firm, management’s list had 8 AI tools; the evidence — expense claims, app approvals, admin settings and a staff survey — showed 17. The inventory is where every later step starts, and the one thing a client or an examiner can ask to see on day one.
What to do this quarter: name one owner for the inventory; list every AI tool with its owner, what data it sees and who pays for it; decide which three use cases matter most and assess those first. Not legal advice. Not a certification.
Sources: MAS Guidelines on AI Risk Management (PDF), paras 1.8, 4.5, 4.12 · TWB sample report (fictional company)
Prices · 18 Nov 2026
AI seat prices did not move this autumn. The plans did.
Our engine reads the public pricing pages of Claude, Microsoft 365 Copilot and Gemini every morning, and ChatGPT’s once a week because OpenAI’s pages refuse automated readers. It keeps every change with its date and source. The calculator on our home page uses those figures, which is why it shows the day each price was checked.
What the record shows for business seats at ChatGPT, Claude, Microsoft 365 Copilot and Gemini between 29 October and 18 November 2026: not one published seat price changed. ChatGPT Business Standard stayed at US$25 a seat a month (US$20 billed annually), Claude Team Standard the same, the Microsoft 365 Copilot Business add-on at US$25.20 (US$21 annually), Google Workspace Business Standard at US$16.80 (US$14 annually). The idea that list prices jump every week is not what our data says.
What did move was everything around the price. Claude retired its “Max 20x” plan on 31 October and listed “Ultra” at the same US$200 the next day. ChatGPT’s page began listing a legacy Codex business seat on 15 November — sold to nobody new, still on some invoices. Gemini Enterprise’s Standard and Plus editions showed a starting price on some days and “contact sales” on others. And the same seat still has two prices, annual and monthly, with the monthly one about a quarter higher; most companies pay monthly and budget annual. The invoice is in US dollars and the budget in Singapore dollars, so the exchange rate alone moves the number.
A budget set once in January therefore drifts by spring — not because a vendor raised a price, but because the plan you bought was renamed, retired or re-tiered, and nobody was reading the page. The fix is not a bigger budget; it is a monthly reconciliation of seats paid, seats used and the plan as the vendor sells it today.
The two-minute version is on our home page: tick the tools you pay for, enter seats paid and seats used, and see what idle seats cost at today’s list prices. Nothing is sent or saved.
Sources: TWB price engine, business plans of ChatGPT, Claude, Microsoft 365 Copilot and Gemini, 29 Oct – 18 Nov 2026: zero plans with more than one published price; each row carries its checked-at date and source page · claude.com/pricing as read by the engine on 30 and 31 Oct 2026 (Max 20x → Ultra)
Vendors · 12 Nov 2026
AI vendors retire models on a schedule. Does your company know which tools depend on them?
Every large AI vendor retires older models on a published schedule: OpenAI keeps a deprecations page that lists each model and the date it stops answering. A tool built on a retired model does not fail loudly — it answers differently, or an integration someone set up two years ago quietly stops.
That is why an AI register is more than a list of subscriptions. For each tool it should say which model or vendor it depends on, who owns it, and who checks it after a change. The HKMA’s November 2026 discussion paper on agentic AI proposes the same for banks: a record per deployment with its stop mechanism, its change history and what happens at retirement — a practice, not a rule.
How to find out what depended on a retired model: start from the vendor’s own retirement list, then look for the model name in your integrations, scripts and the settings of the tools you pay for. If nobody in the company can do that in an afternoon, the register is the gap.
Our AI radar reads vendor announcements twice a day and lists model and plan changes for the tools companies pay for. Not legal advice. Not a certification.
Sources: OpenAI, Deprecations · HKMA, Discussion Paper on Agentic A.I. Governance in the Banking Sector (Nov 2026)