Econify.AI Newsletter, October 16, 2026

JPMorgan is locking AI agents out of customer accounts

JPMorgan said on its earnings call that it will block third-party AI agents from accessing customer accounts. CEO Jamie Dimon warned about thousands of rogue agents and a liability transfer customers don't understand; CFO Jeremy Barnum said the bank is blocking them to protect customers and itself. JPMorgan will still support read-only open-banking access, but it won't refund money an agent spends on your behalf. Shopify went the other direction, opening its API to agents from any provider, including Meta's Muse, Hermes, Instinct, and Dots. The same split is showing up in business software, where Salesforce has opened its APIs to outside agents while SAP has tightened them.

Econify's take: The smart path is to embrace the API and MCP routes and skip the heavy web scraping the alternative demands. How fast you move depends on your business model.

Two new frontier models landed

Google released Gemini 4 Argon and Meta released Muse Wonder and Wonder Full within days of each other, both scoring near the top of public benchmarks. Hands-on, the picture fractured. Argon led several knowledge-work and writing tests but trailed on real coding, and one tester's bug-hunt result swung from mediocre to near-best-in-class simply by changing the tool wrapped around the model. Muse Wonder drew praise for agentic work and sharp criticism on real engineering tasks, where it landed well behind OpenAI's GPT Sol. Meta's Alexander Wang conceded the team hadn't trained for some of the demos people judged it on.

Econify's take: People are sharing just how much better Gemini 4 Argon is without the default harness. Google says doing that violates its Terms of Service. The big win might be Haiku 5.5's release, which is incredibly capable for the cost.

OpenAI's AI solved problems that defined careers, and created a backlog

OpenAI published 722 manuscripts organized into 372 result families, generated by an unreleased internal model, fully solving 90 of the 500 hardest open problems on one widely cited list and making partial progress on two more of the field's hardest problems. Each result reportedly used compute equal to about three hours of ChatGPT Pro use, down from roughly 10,000 agents running 88 hours for a single proof a month earlier. The catch: mathematicians now have to verify and understand hundreds of proofs faster than any human can read them. As one noted, results that aren't understood and connected risk sitting unused.

Econify's take: When AI generates output faster than your people can check it, verification becomes the constraint. Before you point generative tools at high-stakes work, decide who verifies the output and how. If you can't answer that, more volume just means more unchecked work. Overall, this is still an impressive display as to how fast AI is moving.

Other news

Nvidia is in early talks to buy or deepen its stake in Reflection AI, maker of the open Beam model we covered last week, and Amazon is in talks to buy chip-software startup Decart for about $7 billion. Microsoft unveiled its Maia 300 inference chip, claiming 40% more tokens per megawatt than leading GPUs, with deployment starting in 2027. An OpenAI engineer was charged after a recording device turned up hidden in an executive conference room; OpenAI says no model weights, code or customer data were accessed. ChatGPT now answers with interactive charts and diagrams, and Anthropic followed with Claude Dashboards, which build live charts from connected data.

On the money side, OpenAI's annualized revenue is closer to $50 billion than the reported $68 billion, which knocked AI stocks, while SpaceX and Broadcom are each lining up $40 to $50 billion in debt to pay for AI chips. Character.AI filed for bankruptcy, citing costs and litigation, with Disney bidding $20 million for some of its IP. And REA, an open-source tool that lets coding agents reverse-engineer apps down to native binaries, is a reminder that anything you ship client-side can be inspected and rebuilt.