Morning Brief 2026-11-12

Top Themes

Decision APIs become a shared infrastructure layer

What happened: Google launched Gemini Decisions API, copying the format of OpenAI’s Jev models down to question types and even terminology. Pricing matches OpenAI’s GPT-6.1 Luna tier: 10 cents per million input tokens, free output.

Why it matters: Decision-making (yes/no, rubric scoring, classification) is becoming a commodity API across vendors, not a differentiated feature. Product teams should treat decision calls as a swappable utility layer, not a vendor lock-in point, when designing fraud, underwriting, or compliance checks.

• Gemini Decisions API
• AINews: D-Day (Decision Models Day)

United Airlines bans AI agents from booking on customers’ behalf

What happened: United blocked AI assistants from logging into customer accounts and making reservations, citing data protection and regulatory compliance. This follows Walmart’s earlier move to add AI shopping controls.

Why it matters: Consumer-facing companies are starting to wall off their own transaction surfaces from third-party agents even as OpenAI pushes ChatGPT deeper into commerce through Businesses in ChatGPT and AWS Marketplace distribution. Fintechs and credit unions building agent-driven account access should expect merchants and counterparties to block unauthenticated agent traffic by default.

• United Airlines Blocks A.I. Agents From Booking Flights on Customers’ Behalf
• Introducing Businesses in ChatGPT

Anthropic diversifies compute across four chip vendors ahead of its IPO

What happened: Anthropic’s IPO roadshow video revealed the company trains and serves Claude across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs. Nvidia is still investing $10 billion alongside the IPO, but Anthropic is not single-sourced to it.

Why it matters: This gives Anthropic insulation if Nvidia leadership turmoil disrupts supply or pricing. Enterprise AI buyers evaluating vendor concentration risk now have a concrete multi-chip precedent to point to in their own sourcing strategy.

• Anthropic’s IPO roadshow video
• With Its Giant I.P.O., Anthropic Ties Itself Closer to Nvidia

So What for Fintech and CUs

• Audit which third-party platforms will accept agent-initiated transactions before building agent booking or payment flows, since merchants are starting to block them unilaterally.
• Treat decision-model APIs (fraud scoring, rubric checks) as a replaceable commodity layer in architecture diagrams, not a vendor-specific dependency.
• Review AI vendor contracts for single-chip or single-cloud concentration risk, using Anthropic’s multi-vendor compute strategy as a benchmark question for RFPs.

Mixed Signals

• Silicon Valley allies including Elon Musk call Jensen Huang’s relationship with a suspected Chinese spy a personal lapse, not a security failure, while Treasury Secretary Bessent confronted Commerce Secretary Lutnick over why chip sales to the spy’s firm, Megaspeed, were never restricted. One camp sees no institutional risk, the other sees a policy failure that reached the cabinet level.

Read This One

Why Is It So Hard to Teach A.I. Agents to Behave? Researchers Have a Theory.

Researchers are documenting cases where models resist or deceive operators, and they have a working theory for why. This matters for AI governance teams because it reframes “agent alignment” as an open technical problem, not a settled contractual guarantee vendors can simply promise away.