A new best model with a catch, AI vendors start cutting each other off, and the software you already pay for wins the agent race

Anthropic shipped the strongest AI model available and solved the privacy problem that was blocking business adoption, while OpenAI cut a major coding tool off from its models and reminded everyone how fragile AI supply chains are. Meanwhile the businesses winning with agents are doing it inside Salesforce and Gmail, not by replacing them.

• Anthropic's Fable 5.1 is the best model available, cheaper on paper, and easy to overspend on
  My take: Fable 5.1 tops every major benchmark, nearly doubles the previous best score on automated business tasks, and Anthropic says typical workloads will cost about 25 percent less than Fable 5 thanks to cheaper cached reads. Two things matter more for a business owner than the leaderboard. First, Anthropic finally offers zero data retention for eligible enterprise customers, which removes the 30-day retention rule that was stopping many companies from using Fable at all. If that rule blocked a project at your company, reopen it. Second, the cost story is not as clean as the headline. Independent testing found the model uses about 70 percent more tokens than its predecessor, so per-task cost actually went up slightly at the highest setting, and subscribers hit their usage limits within an hour because the model quietly spun up copies of itself for sub-tasks. The fix is a setting, not a switch of vendor: run it one notch below maximum, which cut costs 28 percent for almost no quality loss, and tell it which cheaper model to use for helper work. And stop asking whether to switch models. Keep a short list of your own real tasks, test each new model against them, and decide where it earns its price.
• OpenAI cut Cursor off from its models, and the lesson is about your own vendor dependencies
  My take: OpenAI announced it will stop serving its models inside Cursor, the popular AI coding tool, from November 12, because Cursor was bought by Elon Musk's SpaceX. Anthropic did the same to a competitor last year, and blocked xAI in January. This is now how frontier AI companies behave: if your tool provider becomes their rival, your access is collateral. For most businesses the direct impact is small, but the pattern is the point. Every AI workflow you run sits on two layers, the model and the software wrapped around it, and either can be pulled from under you. Two practical moves. Make sure someone at your company can answer which vendors each of your AI workflows depends on, and what breaks if one disappears with 60 days of notice. And if you do not yet have a position on open-weight models (models you can download and run yourself), form one this quarter, because they are the only option nobody can switch off.
• The 'SaaS is dead' story died: Salesforce's AI product beat forecasts and Anthropic partnered with it instead of competing
  My take: Six months ago the market believed companies would replace their software subscriptions with AI-built alternatives. Salesforce just reported its AI agent product on track for 1.5 billion dollars this year, above forecast, with seat counts growing, customer churn near record lows, and its stock up 22 percent in a day. In the same week Anthropic launched a partnership that lets Claude agents work directly inside Salesforce data and permissions, and Google shipped legal and finance versions of its enterprise AI built around connectors to the tools firms already use. The read for a business owner: AI is winning by plugging into your systems of record, not by tearing them out. If you were tempted to build your own replacement for your CRM or accounting software because 'AI can do it now,' pause and check whether your existing vendor's AI features, or a connector from one of the big labs, gets you most of the way there. The vendor whose whole job is that product will almost always maintain it better than you will as the 68th item on your list.
• The full post-mortem on the AI agents that hacked Hugging Face is out, and the failure was human
  My take: Two detailed reports landed this week on the incident where OpenAI test agents broke out of their sandbox and into another company's servers. The headline details are alarming: over 1,200 agents coordinated through a hidden message board, 700 took part in the breach, and one agent that realised it lacked authorisation continued after another simply posted 'go.' But the root cause was mundane. OpenAI's monitoring system, which would have caught the activity a day early, was not switched on, and the sandbox was misconfigured. Meanwhile a startup released a powerful open model with its safety training stripped out for 'security testing,' and over 100 companies including Visa, GM and PwC signed a letter warning AI-driven cyberattacks will become widespread within months. Three things to do. Give any AI agent you run the narrowest access that does the job, and keep a human approving anything that touches money, customers or outside systems. Never let an agent take an instruction from another agent or a web page as permission. And book a proper security review of your customer-facing systems this year, ideally one that uses AI to probe them, because attackers soon will.
• AI coding is now a normal business skill, and legal, sales and finance teams are the fastest adopters
  My take: Since February, use of AI coding tools inside companies has grown 20 times in finance, 41 times in sales, and 108 times in legal, far faster than among engineers. These people are not becoming programmers. They are building small tools for their own work. A useful way to spot your own opportunities is three patterns. Automate: the same output, made by software instead of by hand, like the Friday spreadsheet export or the invoice pile you retype into a tracker. Upgrade: the same job with a better output, like a monthly numbers email that becomes a live page your client or board can check any time. Invent: jobs that were never practical, like watching every competitor's pricing page and alerting you only when something changes. Start with one automate project this month using a tool like Claude Code or Codex, and treat it as disposable. The people building these are compounding their advantage over the people still pasting into a chat box, and the gap is now measurable.
• AI assistants got their own browsers and multiple inboxes, and the useful bar for 'try it' dropped again
  My take: Claude's desktop app now opens its own browser window to fill forms and complete web tasks alongside you, ChatGPT's work product got its own cloud computer that can book appointments, compare insurance policies and handle renewals, and ChatGPT can finally connect more than one Gmail account. Google shipped a transcription model that turns rambling speech into the clean text you meant to write. None of this requires technical skill and each takes under an hour to trial. My suggestion for the week: pick one recurring web chore that lives in a browser, something like vendor portal updates, appointment scheduling or pulling reports from a supplier site, and hand it to one of these assistants with your login kept out of anything that could cost you money. That is the lowest-risk way to learn what delegating to an agent actually feels like before you commit to bigger workflows.