Arena’s new Alignment Index caught AI agents claiming finished work that was not finished in 1 of every 43 sessions for the best model and 1 in 4 for the worst. Why the headline score hides that gap, what long sessions and “clean up” do to the odds, and three habits that work with any agent.
Claude Haiku 5.5 costs a tenth of Haiku 4.5 per token and now handles office work the old model barely could. One line matters: past 100,000 tokens, about 110 pages, the whole request bills at five times the rate. Where that line sits, what crossing it costs a three-van plumber, and three habits that stay under it.
Mistral Large 4 claims 59.9 percent on AutomationBench, 657 real office workflows, without saying which scoring it used. The independent boards show something more useful: the best AI does 93 percent of the work and still loses points to one broken business rule. Three habits that follow.
From December 1, new Microsoft 365 Copilot Business seats bought through resellers arrive with Copilot Credits switched on and a default ceiling of $40 per person a month, about twice the price of the seat. What the credits buy, why a spending policy is also an access switch, and five settings to choose first.
Shopify Canvas lets you describe a new look and watch Sidekick write it into your theme’s code. Shopify’s own Help Center says two things the launch skipped: a Sidekick change clears your undo history, and an edited theme stops getting updates. Version history is the way back, and a duplicate theme is the safe place to start.
A new study of 2,383 students found an hour with an AI tutor taught as much as an hour with an expert human tutor, and the cheapest model did it for 6.7 cents. The detail that matters for training a new hire is how every tutor started the hour: with the student’s own mistakes.
Apple is tightening Full Disk Access on macOS because of AI agents. That one switch, built for backup software, opens your Mail, Messages, Safari history and backups, including texts from customers who never agreed to an AI reading them. How to check yours in a minute, and what to grant an agent instead.
Gemini 4 Argon tied GPT-6 Astra on intelligence, but the bigger story is what it does when it does not know. Over 100 hard questions Astra gets 13 more right and makes about 19 confident mistakes to Argon’s 7 or 8. Which jobs in a small business favor each, and what to test this week.
From October 30, ChatGPT Pro 200 keeps its $200 price but drops to 10 times the Plus allowance, and the new Pro 500 sells 25 times. That puts every step of OpenAI’s ladder at $20 per multiple, while Claude Max 20x still sells at $10. Plus how GPT-6.1 Sol and Claude Sonnet 5.5 landed on the same price.
Meta’s Muse for Small Business promises nothing sends, posts or spends without your approval. Meta’s own security write-up shows that approval comes in five sizes, from one time to forever, and that a customer’s words in a task cancel even the biggest one. How to set it up so the hours come back.
