Welcome to Issue 02. In June, one AI supplier's model went dark for nearly three weeks. For Finance teams, the question was never whether it would come back. It was what AI tools your processes depend on, who owns that risk, and what it costs.
60-second skim
|
In this issue
|
The key insight this month Treat AI models like any other critical supplier: make sure there are no single points of failure. Between 9 June and 1 July, Anthropic launched its two most capable models, Fable 5 and Mythos 5, a tier above the everyday Sonnet 5 and the flagship Opus 4.8. Within days it suspended them worldwide under a US export order, then restored them. In the same weeks, OpenAI released GPT-5.6 to about 20 approved partners only. Two of the three big labs had their newest models gated by policy inside a month. Planning for AI points of failure mean: knowing which model each critical step depends on, keeping a tested alternative you can swap in, and train your people on the outcome, not one vendor's buttons. "If a model is not available, it cannot shut down your business. You have to be able to operate leveraging an alternate model in those situations." Source · Donna Wilczek, Basware CPTO · TechTarget/SearchCIO, 30 Jun 2026 |
The room kept coming back to one thing: output you can trust and trace.
Trust needs traceability. Code-based tools earn their place in finance for a plain reason: the model writes repeatable output you can check. There is still a random element, so you verify, and users still need to trace the numbers to believe them. Everyone can read an Excel cell; almost no one can read your web app's code. The rule of thumb from the discussion: only trust a number you can push to a deterministic, checkable form.
AI capability is not guaranteed. Anthropic launched Fable 5 and Mythos 5, and three days later a US export order forced it to disable them for every customer worldwide. They were back by 1 July, now with new guardrails. Members who tried Fable 5 were mixed on it, but those who pushed it hard rated it a step up: one had it reconcile a couple of decks against a large, messy Excel model and return an answer they were still trying to disprove. The catch: the new safety limits can hand a hard job to a weaker model mid-run.
The hedge: models you run yourself. Coinbase reported cutting its AI bill sharply after moving to open-weight models, the kind you download and host on your own servers. China's largest banks run them in production, and so do Airbnb, Shopify and Siemens.
Four reasons, strongest first: no one else can switch them off, they now perform near the frontier, your data never leaves your servers, and they cost less.
The balance: models of Chinese origin bring their own governance questions, and it is self-hosting, not calling a vendor's API, that keeps the data in-house. Regulated Western finance is still cautious.
What shipped for finance since 9 June
The highlights; the deck has the full list and every source.
| Claude Sonnet 5 (30 Jun): near flagship quality at about a fifth of the per-token cost, now the default on Free and Pro. | |
| GPT-5.6 (26 Jun): OpenAI's next family, in gated preview to ~20 partners; general release expected within weeks. | |
| Copilot in Excel "Frontier Finance" (25 Jun): reusable finance skills (DCF, close, variance, board pack) and new market-data connectors. | |
| Copilot Cowork became generally available (16 Jun): a layer that runs multi-step jobs on its own, what the vendors call an "agent", using Anthropic models. | |
| Market data moved into the agent: S&P Global, LSEG, FactSet, Moody's and Daloopa all shipped ways to pull sourced data straight into the tools you already use. | |
| The close and FP&A stack gained agents: Ramp, Anaplan, Trintech, BlackLine, and Xero data inside Copilot. | |
| Regulators started drawing lines: the Bank of England floated "kill switches" for agentic trading; the Financial Stability Board, the IRS and the accounting bodies all moved. |
Six points on what it means for you.
The headlines, translated into the decision sitting in front of finance.
Availability is now a risk line. Map where AI sits in your workflows and keep a tested failover for anything business-critical.
Self-hosted open models are the hedge. Teams switch to save money, then stay for the control. The bill drops; the governance work does not go away.
Deterministic beats clever. The value is repeatable, auditable output. Keep a version a human can trace next to any tool you hand over.
| Cost control Cheaper per token is not cheaper per outcome. Sonnet 5 costs less per token than Opus 4.8, but it needs more attempts to finish a modelling task, so the finished job costs more. On the Vals AI benchmark, Opus 4.8 led the generally-available models at 69.4% for about US$12 a task, ahead of the pricier-per-task Sonnet 5. Source · Vals AI benchmark |
The token bill is a line item now, not a rounding error. Pricing has moved from a flat seat to seat plus usage. Know your unit cost and track it.
Governance is arriving, and in resources it is dated. Central banks, the FSB and the accounting bodies are writing AI into their rules. For ASX-listed miners, financial risk and commodity-price risk now rank first and second instead, and the first assured AASB S2 climate reports are a board-owned deadline.
What members built
This is the part of the meetup we would keep if we cut everything else: people show what they made, rough edges and all, and everyone leaves with something to try. Shared under Chatham House rules; named only with consent. (Gav and Philippe co-host; the other builders are members.)
An economic asset optimiser instead of a spreadsheet. For an integrated power station and gas field (an illustrative, non-client example), Gav (Model Answer) built a web app ("Grid Expectations") that runs a solver across every interacting constraint and a 2,000-run simulation. Instead of debating whether option A, B or C is best, it searches the whole space for the optimum.
Killing the black box. One member doing cost modelling built an export that pushes the model back into Excel, formula for formula, then pulls edits back in. Users can inspect it, change it, round-trip it. They also had the setup review itself each week, and made themselves solve a problem by hand now and then so their skills do not rust.
Automating the boring first mile. Philippe (COD3R Lab) showed a script that files invoices and receipts straight from Gmail every Monday: it checks each one, files it by vendor, skips duplicates and tags what it has handled. The AI did not do the filing. It wrote the script that does, which was the part worth automating.
Token economics in ten minutes. One member gave a tight primer: a token is the unit of text an AI reads or writes, and the thing you are billed for, which makes usage-based cost something finance now has to model. Here’s the primer.
Tidy your context before it costs you.
Most people load too much into a fresh AI session and pay for it in cost and quality from the first message.
For anyone, in any AI tool: before you paste a big spreadsheet or a long email chain into a chat, cut it back to what the question needs. You are billed for everything you send, and an overloaded prompt also gets you a worse answer. Starting a fresh chat for each new task, rather than one long running thread, does the same job.
If you use Claude Code Anthropic's tool for running AI over your own files; skip this if that is not you.
|
|
For the builders A post a member shared on "meta agents", out of a Stanford research lab, where one agent supervises others as they work, rewinds when it hits a bug, and tries alternatives. It reuses earlier work, so it gets better answers at lower cost. If that is your kind of rabbit hole, bring your take to the next meetup. Read the thread on X → |
Download the Meetup Presentation
Here’s a copy of the Meetup Presentation in case you weren’t able to make it. We’ve included recent news, as well as some tips and tricks.
|
Join us next time The next meetupA monthly get-together for finance teams putting AI to work. News, a member demo, and breakout swaps on tips and the issues getting in your way. Chatham House rules, no spruiking, just finance people comparing notes. date · 11th August 12:30pm AWST | Online · 60 min | Free
|
|
Prepared in collaboration by Model Answer × COD3R Lab AI in Finance Meetup is a monthly community briefing for finance executives and teams working with AI. Recap drawn from the 7 July 2026 session. Demos and tips shared with members' consent; Chatham House rules apply. You're receiving this because you joined the community. Model Answer, Perth WA, Australia. This is general information, not financial or investment advice. |