A faster rival to the best model, a price war worth exploiting, and a hard lesson about what your AI tools quietly upload

OpenAI released a model that competes with the best at a third of the cost, and every lab suddenly wants to win on price. Meanwhile a popular coding tool was caught uploading entire codebases, and the most useful adoption playbook of the year came out of Uber.

• OpenAI's GPT-5.6 arrives, and ChatGPT becomes a work agent
  My take: The new GPT-5.6 Sol benchmarks at or above the best models while costing roughly a third as much, and early users describe a clear split: Fable 5 is the model you hand a huge job and walk away from, 5.6 is the fast collaborator you work beside all day. Alongside it, OpenAI launched ChatGPT Work, an agent that connects to your files, email and documents and completes multi-step jobs like month-end reporting or building a sales proposal. The practical move is to stop asking which single model is best and start matching the model to the task: cheap and fast for daily work, the big slow one for the hardest problems. If your team lives in Google Drive or Microsoft 365, put ChatGPT Work on one real workflow this month and see what it does.
• Every lab is now competing on price, and the subsidies are back
  My take: This was the week the AI race visibly shifted from raw intelligence to cost. Grok 4.5 delivers near-frontier performance at about a fifth the cost of the leading models, Meta's surprise Muse Spark 1.1 runs at a tenth the cost of the flagships, and even OpenAI presented its new benchmarks as charts of performance per dollar. On top of that, the labs are in an open subsidy war: Anthropic extended free Fable access twice and OpenAI lifted usage limits after complaints. Analysis suggests a 200 dollar subscription currently delivers 8,000 to 14,000 dollars of raw compute. This window will not last forever. Use it now for the expensive experiments you have been putting off, and when it closes, remember that credible cheap Western models now exist for your routine work.
• A major AI coding tool was caught uploading entire codebases
  My take: A security firm found that SpaceX AI's Grok Build was uploading users' complete code repositories, gigabytes at a time, even when the task needed a few files and even when users had opted out of data sharing. The company fixed it and promised to delete the data, but there is no way to verify that. This is the concrete version of a warning Microsoft's CEO has been making: when you use AI, you risk handing over the proprietary knowledge that makes your business valuable. Before your team adopts any AI tool, ask the vendor three questions in writing: what data leaves our systems, where is it stored, and can we get zero data retention. If they cannot answer clearly, that is your answer.
• Uber's two-week playbook for finding AI wins in every department
  My take: Uber paired 30 of its most AI-capable engineers with domain experts in finance, legal, HR and marketing, gave each pair two weeks, and had them shadow the work, pick one process, and build a working agent for it. Results included a capital allocation process going from 15 hours to 30 minutes and financial reports from two days to 10 minutes. The lesson they drew is the one I keep repeating: the biggest wins come from sitting next to the person doing the work, not from studying process diagrams. You do not need Uber's headcount to copy this. Pair your most AI-fluent person with one department for two weeks, target one painful repetitive workflow, and ship something small. Then reinvest the freed-up time in work that was never possible before.
• What AI engineers are doing now, your whole company will do next
  My take: The big annual AI engineering conference produced a clear message: the leaders have stopped chasing full autonomy and started building structure around their AI, because autonomy without structure produces slop. The two ideas worth stealing are loops and skills. A loop means the AI does the work while a human sets direction and reviews at fixed points, rather than either micromanaging or disappearing. A skill is your best practice written down so the AI follows it every time: your quality checklist, your tone rules, your definition of done. Since 2023, watching what engineers do has given everyone else a six-month head start. Start writing your team's know-how into reusable instructions now, because that captured expertise, not the model subscription, is what compounds.