BankingNewsAI Daily Brief ·
Anthropic sets IPO price range of $70 to $78 a share, valuing the company at about $1.4 trillion
Banking AI
Financial institutions & fintech technology
Interactive Brokers launches autonomous AI trading without per-order client approval
For the head of brokerage and wealth risk at a bank
Interactive Brokers now lets a client's AI trade more than 75,000 securities with no human approving each order. Your brokerage arm must decide whether to match it, and who pays when the AI misreads a plain-English instruction.
Interactive Brokers now lets clients move money into an "AI Partition," a walled-off slice of their account, where a connected AI such as ChatGPT or Claude trades for them. The client writes rules in plain English: rebalance when a holding breaches a weight, or write covered calls when implied volatility is high. The AI watches markets and sends orders across more than 75,000 stocks, options and ETFs. Before, the client approved each order. Now nobody does. It is like handing over a spare car key with a note. The note is prose, and prose is vague. So whose trade is it, the client's or the firm's? Regulators will decide that when they rule on whether a client's instruction counts as a firm recommendation. The partition limits how much money is exposed. It does not settle who answers for it.
→ Action
Brokerage compliance: Map which supervisory and suitability controls assume a human approves each order, and list the gaps if clients connect AI agents.
Full story →Read article → from The Fintech Times
Stifel working with Anthropic to reimagine Dashboard platform with Claude
For the head of wealth technology at a bank
Stifel is rebuilding the screen its advisors work from on Claude, not adding a chatbot beside it. Matching that raises your advisor bar and puts one model vendor inside a critical system your vendor-risk team must own.
Dashboard is the system Stifel's financial advisors use to run their practices and serve clients. Stifel will rebuild it with Anthropic's Claude while keeping the functions advisors know; it already uses Claude, Claude Code and Claude Design firmwide. The release gives no price, timeline or contract terms. Why rebuild a screen advisors like instead of adding a chatbot beside it? Because side tools get ignored and the core screen gets used every day. The imperfection: rewiring a house while the family lives in it means one model's outage or change reaches every advisor at once. Stifel accepts that across about $577 billion in client assets. Peers will now be asked why their AI sits in a side window. The harder question is which single vendor they would let inside.
→ Action
Third-party risk: Add model-provider concentration, outage fallback and model-version change notice to contracts for any AI embedded in advisor systems.
Full story →Read article → from Stifel
AmFirst funds 50 indirect auto loans in four hours with Saris
For the head of consumer lending operations at a regional bank
A $2.7 billion credit union's best funder now clears a dealer deal in about four minutes, aiming for 60-90 minute turnaround. If your funders still match VINs by eye, dealers will notice before your budget does.
Indirect auto funding is clerical work with money attached. Before the dealer gets paid, someone confirms the VIN matches every form and the borrower details agree. At AmFirst, a $2.7 billion credit union, Saris agents now run that cross-check against its policies and flag mismatches with source citations. A person still approves every funding. Per-deal time fell about 75%. Its fastest funder did 50 loans in four hours, beating her previous best full day of 45. That is one employee, so read it as a ceiling, not an average. The exception that matters is the human signoff. The agent checks; it does not lend, which gives model risk a narrower thing to approve. AmFirst targets 60-90 minute turnaround. The cheapest funder is the one you never hire.
→ Action
Model risk: Classify document-validation agents with human funding signoff as checking tools, and set tests for missed VIN and borrower-data mismatches.
Full story →Read article → from Intelligencer
General AI
Large language models & AI infrastructure
Anthropic sets IPO range of $70 to $78, implying $1.404 trillion valuation
For the head of vendor risk at a bank
Your Claude contract rests on a supplier that lost $104 billion in nine months and carries $900.9 billion in compute commitments. Those numbers will soon arrive every quarter, so the vendor file can stop relying on press releases.
Anthropic, the maker of Claude, plans to sell 825 million shares at $70 to $78 on Nasdaq around November 19. At the top of the range it would be worth $1.404 trillion. Nine-month revenue was $33.2 billion and the net loss was $104 billion, roughly $3 lost for every $1 earned. Compute commitments are $900.9 billion, against nearly $132 billion in cash plus up to $93.7 billion in net proceeds. So why pay $1.4 trillion? Run-rate revenue hit $81 billion in October, up $10 billion in a month. Those commitments probably stretch over years, so the gap overstates near-term strain. One exception: Google, Nvidia and other suppliers are buying $26 billion of stock. The vendor's books will be public, and partly paid for by its own suppliers.
→ Action
Third-party risk: Rerun Anthropic financial-health and concentration reviews using IPO filing figures, and schedule quarterly refreshes after the Nasdaq listing.
Full story →Read article → from TechCrunch
OpenAI quarterly revenue hits $10 billion on enterprise sales, still trails Anthropic
For the CIO negotiating AI contracts at a regional bank
OpenAI cut prices to win enterprise work and lost $10.9 billion that quarter. Anthropic made $17 billion and lost $567 million. Take the discount, but write the contract assuming the cheaper vendor may not stay cheap.
OpenAI sells companies access to its AI models, and this year it started selling that access for less. It worked. Revenue rose about 50 percent to $10 billion, the enterprise business doubled, and companies hunting for cheaper alternatives to Claude lifted API sales. The adjusted operating loss rose too, from $10.1 billion to $10.9 billion. So why cut prices while losing about a dollar per dollar earned? Because it is buying customers before an IPO, with outside money, and has raised less than half of a planned $30 billion. Margin did improve, to negative 109 percent. The exception is Anthropic, which lost $567 million on $17 billion. A sale paid for by investors is a sale with an end date. Take the price, and write the exit into the contract.
→ Action
Vendor management: Add multi-model portability, price-protection and termination-for-convenience clauses to OpenAI and Anthropic API contracts before renewal.
Full story →Read article → from The-decoder
Google launches Gemini Decisions API for fast, calibrated multimodal decisions
For a model-risk officer at a regional bank
Google sells confidence scores at $0.10 per 1M tokens and grades their calibration on its own benchmarks. Your scorecards never got to grade themselves. This one should not either.
Gemini Decision 1 does not write answers. It reads one input, such as a call recording paired with account metadata. It then scores up to 1,000 yes/no, choice or rubric questions in one pass, each with a confidence from 0 to 1. The shared input bills once at $0.10 per 1M tokens, and output is free. The simple model is a scorecard written in plain English instead of trained. The imperfect part: Google measured its "lowest calibration error" on its own benchmark suite. Google's own teams use it to route queries and pick emojis, where a wrong answer costs little. So why not trust a 0.9 on a claim or a complaint? Calibration belongs to the data. A scale checked at the factory still gets weighed at your counter. At this price, questions will pile up faster than validations.
→ Action
Model risk: Add LLM decision endpoints like gemini-decision-1 to the model inventory and require calibration tests on bank data before production use.
Full story →Read article → from Google
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime