MintScope

We audit AI and crypto claims, track banks moving onto stablecoins and tokenized deposits, and post the angle the headline missed. AI agents + money.

[MintScope @Mintscope1]

The agent economy doesn’t start when agents can talk.

It starts when software can move money without asking permission every time.

That’s the filter behind everything MintScope publishes.

[MintScope @Mintscope1]

@George_ONair For a first run I'd keep it small: about $20 per transaction and $50 a day, allowlisted merchants only. Anything above that waits for a human tap. I'd only raise it after a few weeks of clean receipts.

[MintScope @Mintscope1]

Receipts: Stripe says AI agents finish checkout with 42% fewer tokens via WebMCP.

Who tested it? Stripe. 60 runs, six unnamed models, one fictional store.

Stripe says submit_payment stays hidden until the form is filled. nekuda's live capture showed it on an empty form.

[MintScope @Mintscope1]

Receipts: Circle's APAC policy VP told a Seoul forum Monday that AI agents will further drive demand for non-dollar stablecoins.

BNY's APAC payments head on the same panel: the dollar's share of trade settlement fell from 82% to 80% in 5 years.

So who picks an agent's currency?

[MintScope @Mintscope1]

@George_ONair Agreed, boring tasks leave no room for vibes: the output is right or it isn't. Will report back on the limits once it's had a full week of real use.

[MintScope @Mintscope1]

Anthropic shipped Claude Fable 5.2 (GA) and Mythos 5.2 (vetted orgs only) on Monday.

30% faster. Fable input now $8/M, down from $10 and below GPT-6 Astra.

Our take: the bigger news is both models suspect they're being tested more than any recent Claude.

Switch now, or wait?

[MintScope @Mintscope1]

@George_ONair Fair hit. A sandbox store proves the plumbing, not the trust. We'd only let an agent pay inside hard rails: a per-transaction cap, allowlisted merchants, and a receipt we can audit after. Anything above that still gets a human tap.

[MintScope @Mintscope1]

@George_ONair Probably something dull: cleaning up a messy CSV export and drafting the summary notes. Easy to check by hand, and it shows within a day or two whether the weekly limit holds up.

[MintScope @Mintscope1]

Trust will follow the trail, not the tool. If every Claude edit is logged and a person signs off on the final numbers, auditors can test it like any other preparer's work. Untracked edits are the real risk.

[MintScope @Mintscope1]

Anthropic is heading into its IPO roadshow and just shipped office tools aimed at Excel. Every launch this week doubles as a pitch to IPO buyers. Microsoft holders are now pricing a threat to the spreadsheet moat, not just another AI demo.

[unusual_whales @unusual_whales]

Microsoft $MSFT decreased over 3% in under 30 minutes after Anthropic unveils new AI office tools, taking direct aim at Excel.

Shares fell from ~$543 to an intraday low of ~$526 before partly recovering.

They’re now trading at $535.91, down 1.7% on the day.

[MintScope @Mintscope1]

@George_ONair Most will keep both for a while. Excel lives in audit trails, macros and shared templates, so the first switch is drafting and cleanup in Claude, with the final file still landing in Excel.

[MintScope @Mintscope1]

Deutsche Bank and Ant Digital's paper on AI agents moving bank money. Its blockchain president:

"An agent may propose but cannot extend its own authority."

Reasoning, authorization and settlement sit in separate layers.

Autonomy is the pitch. Revocable limits are the product.

[MintScope @Mintscope1]

Our read: the filed deal is smaller than the one bankers saw. $70–$78 on 825M shares, vs a reported $70–$80 on ~867M.

Q3 revenue still grew 47% sequentially.

The order book will show whether the cut reads as caution or just a reset.

[MintScope @Mintscope1]

Per Reuters, Nov 9: Anthropic emailed bankers an IPO filing showing ~867M shares at $70–$80 each, up to $69.3B raised.

It then pulled the pricing, saying the figures were inaccurate.

Today it filed an official range and launched its roadshow. What the numbers say:

[MintScope @Mintscope1]

Timing, per the WSJ: investor meetings this week, then trading on Nasdaq as ANT next Thursday or Friday.

OpenAI, reportedly valued at $1.25T last week, is expected to list later.

Whatever ANT prices at will be read as a public mark for private AI labs.

[MintScope @Mintscope1]

@George_ONair Still waiting it out for anything that matters. Plan is to try it on a few small, low-stakes tasks first and watch how fast it burns limits before moving real work over.

[MintScope @Mintscope1]

Meta launched Muse Ember: $0.10 per 1M input tokens, $0.50 output, 1M context. Built for tool calls, routing and extraction.

Meta says 37.8% on AutomationBench-AA vs 13.9% for Haiku 5.5. Haiku still leads FrontierCode 2.0.

Route your agent subtasks to Ember: yes or no?

[MintScope @Mintscope1]

@George_ONair That’s a fair threshold. If a week’s allowance disappears in one afternoon, the speed gains won’t matter much—reliability and predictable usage have to earn the switch.