BankingNewsAI Daily Brief ·
OpenAI launches GPT-6 Sol and Luna
Banking AI
Financial institutions & fintech technology
Six banks publish agentic commerce principles
For the card-payments head at a regional bank
Agent-initiated payments will make proof of customer intent and agent authority a baseline control, not a product feature. Your bank must decide whether its payment records can support disputes, fraud cases and liability decisions when an agent acts for a customer.
ASB Bank, Bank of America, Capital One, Commonwealth Bank of Australia, ING Group and NatWest Group published joint principles for trusted agentic commerce. The principles cover transparency, safety, privacy and data, choice, and interoperability. The banks say the framework is meant to help customers and merchants retain choice, control and flexibility over payments while keeping payments safe and secure. Bank of America says confidence will require approaches to identity, authorization, fraud prevention, liability management and customer protection. The consortium plans a later paper on how to apply the principles and invites banks and other payments institutions to help develop it.
→ Action
Payments authentication: Check controls for customer authorization and agent identity in agent-initiated payments.
Read article → from Bankofamerica
Schwab and Anthropic block model training in Claude deal
For the wealth-platform leader at a regional bank
Schwab has set a contract floor for adviser AI: no model training and less data exposure cannot depend on an enterprise license. Your vendor terms must cover every adviser tier before AI reaches client data.
Schwab integrated Anthropic’s Claude for Advisors across its custodian network of more than 16,000 RIAs under terms that bar Anthropic from retaining input or output for model training. Anthropic’s enterprise plan already defaults to no model training, and Schwab extended that protection to lower subscription tiers. Schwab will mask client account numbers and will not pass Social Security numbers or dates of birth to Anthropic. The read-only integration supports meeting preparation, financial-plan updates, analytics, client follow-ups, RMD identification and account-balance reviews. Schwab says it will first assess demand and usage before allowing agent actions with human involvement.
→ Action
Vendor management: Check adviser-AI contracts for no-training terms across all license tiers and for limits on account, Social Security number and birth-date sharing.
Read article → from Investmentnews
Promenaut goes live at HSBC
For a bank operations and model-risk executive
HSBC has put agent orchestration into production for selected workflows, using a platform that supports permissions, audit trails and human review. Your bank must decide whether its controls can prove who authorized each agent action before expanding AI beyond pilots.
Promenaut has gone into production at HSBC for selected operational workflows. The US-based platform orchestrates agents for business case evaluation and pre-coding in the software delivery lifecycle. It supports permissions, audit trails and human-in-the-loop controls for authorizing, executing and reviewing agent activity. HSBC invested in Promenaut, which has raised $8.9 million to date, and HSBC chief AI officer David Rice will join Promenaut’s board. Rice said HSBC wants AI-enabled work to be controlled, accountable and measurable, with clear human oversight and authority.
→ Action
Model risk: Test whether agent workflows log each authorization, action, reviewer and authority level from request through completion.
Read article → from Finextra
General AI
Large language models & AI infrastructure
Anthropic introduces Claude Opus 5.5
For the CIO of a regional bank
Do not carry Opus 5’s cost and control assumptions into agent design. The lower claimed run cost can widen use cases, but stronger resistance to prompt injection does not remove the need to test bank-specific agent controls.
Anthropic introduced Claude Opus 5.5, the first model in its Claude 5.5 family. Anthropic says it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads. Input tokens cost $4 per million, output tokens cost $20 per million, and cache reads cost $0.20 per million; Anthropic says cache reads make up most agentic and coding costs. The company says the model generates output more than 30% faster than Opus 5 and is more resistant to prompt injection. Anthropic says it is less likely to take hard-to-reverse actions or act outside given boundaries, though it still has limits.
→ Action
AI governance: Test Opus 5.5 against prompt-injection and irreversible-action controls before use in agent workflows.
Read article → from Anthropic
OpenAI launches GPT-6 Sol and Luna
For the AI platform owner at a US bank
OpenAI has cut the cost threshold for scaled agent use while splitting models by workload. Re-benchmark routing and control performance before lower prices turn pilots into production demand.
OpenAI launched GPT-6 Sol and GPT-6 Luna in its API, Codex and ChatGPT. The company says the models bring much of GPT-6 Astra’s strengths to faster, lower-cost work at scale. API prices for Sol and Luna are 50% below GPT-5.6 promotional pricing, and OpenAI says it also improved caching and inference efficiency. The company says both models improve on their GPT-5.6 counterparts in alignment. Sol and Luna are available today to ChatGPT Work and Codex users on Plus, Pro, Business, Enterprise and Edu plans, while Luna is also available to Free and Go desktop users.
→ Action
AI platform: Re-benchmark model routing, cost per task and control performance for high-volume and complex agent workloads.
Read article → from OpenAI
Get this in your inbox every morning
Free · No spam · Unsubscribe anytime