Google and Meta are back at the frontier, the cheapest good AI is now American, and banks and Shopify picked opposite sides on agents

Google finally released Gemini 4 Argon and Meta launched Muse Wonder, so there are now five serious US model makers, while Anthropic's new Haiku and Grokbot's switch to using rival models showed price and choice matter more than who is smartest. Underneath, JP Morgan locked agents out as Shopify opened to all of them, OpenAI's real revenue came in 20 billion dollars below the rumour, and the cheapest lever for most businesses is still the person on the team who is already good at AI.

• Gemini 4 Argon and Meta's Muse Wonder both arrived, and neither is a reason to switch your main tool
  My take: Google's Argon is now usable and scores at the frontier, especially on legal and finance tasks, but real users found it weak at working inside existing code and much better when run in Google's paid Ultra tier than in the cheaper plan. The clearest early win is writing: one team that tested it for a week called it their new favourite editing partner, because it structures arguments and explanations better than the rivals. Meta's Muse Wonder also benchmarks near the top and does well on agent tasks, but testers split hard on coding quality, and several burned through a 50 dollar monthly plan in under an hour. The question is no longer which model is best. It is which model is best for which job. If you write a lot of proposals, reports or client letters, give Argon a week as your editor. Keep your main coding and analysis work where it is, and watch the usage meter closely on any new plan in its first week.
• Anthropic's Haiku 5.5 costs a quarter of Sonnet, Grokbot will now use rival models, and Claude can build live dashboards from your data
  My take: Haiku 5.5 is Anthropic's new small model, priced at a quarter of Sonnet 5.5 and half the old Haiku, with an 80 percent discount on short prompts, and it beats OpenAI's equivalent on almost every test. With it, the cheapest capable models in the world are now American rather than Chinese. In the same week Elon Musk said Grokbot will route tasks to Claude, Midjourney and others when they do the job better, and OpenAI is already selling other companies' open models through its own business marketplace. The lesson is that even the labs admit no single model wins everything. For a business this means two things. First, look at any routine high-volume job, such as sorting emails, tagging leads or summarising tickets, and price it on Haiku, because it is now cheap enough to run on everything. Second, Anthropic released Claude Dashboards, which connects to your CRM or accounting software and builds a chart view that updates itself and answers plain-English questions. If someone on your team spends Monday mornings building a report by hand, try this first.
• JP Morgan is blocking outside agents from customer accounts while Shopify opened every store to any agent
  My take: Jamie Dimon said the bank will not let personal agents log in with a customer's password, and that it will not refund losses caused by an agent the way it refunds fraud. Read-only access through approved open banking links will be allowed, so budgeting and finding old subscriptions will work, but paying bills through an agent will not. Shopify went the other way and now lets agents from any provider shop through a direct connection to its stores, after a month of doing it with Meta's Muse alone. Salesforce has opened up to outside agents, SAP has closed down. Two things to do. If you or your staff have given a personal agent your bank login, take it back now, because the liability for its mistakes is yours. And decide which side your own business is on. If you sell things, being easy for an agent to buy from is the Shopify bet, and on current evidence it is the better one for anyone without Amazon's market power.
• The single most reliable way to improve how a team uses AI is a proper champions program, and most companies set it up wrong
  My take: Every team has one person who tries every new tool and whom everyone else asks. Turning that person into an official champion, with time to help others, is the highest-return AI investment I see in companies of any size. It fails in predictable ways. Senior people nominate themselves, the best candidate is too busy to do it, the role is handed out as a reward to someone who is not actually good with AI, or nobody owns it after the launch day excitement. Do it this way instead. Write down whether the role is to spot good use cases and pass them on, or to actually build. Pick someone who already uses AI without being told, whom peers already ask, and whose manager will give them one or two days a month. Give them one test: take one real task from the team and show it working by the end of their training. Then meet as a group regularly, and let people step out when it is not for them. In a small business that might be one person and a monthly half hour, and that is enough.
• OpenAI released 372 new maths results from an unreleased model, and the lesson is which kinds of work fall first
  My take: An internal OpenAI model produced hundreds of new mathematical proofs, including partial progress on two of the most famous unsolved problems, at an average of about three hours of compute each. A month ago one result took 10,000 agents and 88 hours. Mathematicians are now worried less about being replaced and more about having too many results to read and check. You do not need to care about maths to take something from this. AI moves fastest on work where the answer can be checked: maths, code, reconciliations, compliance against a rulebook. It moves slowest where the answer depends on messy information, human judgment and trust, such as medicine, law, leadership and sales. Go through your own business and sort tasks that way. The checkable ones are where you should expect AI to do most of the work within a year, so start there. The judgment ones are where your people's time is worth more, so move them toward it.
• OpenAI's revenue is 50 billion dollars, not the 68 billion that was reported, and Character AI went bankrupt spending 10 million a month on 2.8 million of revenue
  My take: The bigger number came from an investor who counted OpenAI's revenue the way Anthropic counts its own, including the share partners like Amazon and Microsoft keep. OpenAI still grew 77 percent in a quarter, but chip and cloud stocks fell sharply on the news, and Anthropic's IPO filing will face the same accounting question. Separately, Character AI filed for bankruptcy after its compute bill ran at more than three times its subscription revenue, and Disney is buying the pieces for 20 million dollars. Neither story means AI is a bubble about to pop. They do mean the prices and companies you rely on are not fixed. A tool that loses money on every user will either raise prices, cut features or disappear. Before you build a workflow on a small AI vendor, ask how they make money, and keep your prompts, data and process notes in your own files so you can move if they go.