The best model is not the one you need

Anthropic just passed OpenAI on revenue for the first time, and in the same fortnight its most expensive model was reported as the one people are quietly not buying. That is not a contradiction, it is the market growing up. The leaderboard and the invoice have come apart, so the useful question is no longer "which AI is smartest" but "which one is worth what it costs on this job".

This week in AI

Anthropic overtook OpenAI on revenue, while its priciest model underperforms Anthropic took in 11.6 billion dollars last quarter against OpenAI's 6.7 billion, its first quarter in front. Yet the same fortnight brought reporting that its top-end model is losing users to cheaper tools, and that Anthropic tokens cost roughly 4.4 times the platform average on Vercel. Why it matters: being strongest at the frontier and being right for your workload are two different things. Choose the model per task, and re-test that choice every quarter. (qz.com)

Claude agents now run inside Azure, with your data residency intact Microsoft Foundry added five Claude capabilities: structured outputs, web search with citations, web fetch, a connector for MCP servers (Jira, ServiceNow and similar) and dynamic tool loading. All of it runs on Azure-hosted deployments. Why it matters: building a genuinely agentic assistant used to mean giving up European data residency. For anyone already on Microsoft 365 that trade-off is gone, which removes the most common blocker to a first real agent project. (devblogs.microsoft.com)

Both big labs now let your data stay on your side OpenAI launched zero data retention for its frontier models. Anthropic, after enterprise pushback, is moving its 30-day safety retention off its own servers into the customer's cloud, built with over 100 regulated-industry customers and shipping this autumn. Why it matters: "we cannot use American AI because of our data" is now a contract question rather than a technical dead end. If that objection killed a project last year, reopen it. (openai.com)

AI is now its own biggest customer, and unit costs stopped falling On the OpenRouter marketplace, tokens consumed by agents grew fourteenfold between February and August, while human use grew less than threefold. Meanwhile memory prices are up roughly 500 percent in a year and Nvidia has warned of AI server prices about 15 percent higher. Why it matters: an agent working in the background bills nothing like a chat window. Budget agent projects on volume rather than seats, and insist on prompt caching, which covers about 70 percent of agent tokens. (the-decoder.com)

The cheap tier got genuinely good Google shipped Gemini 3.7 Flash with better coding and half the price of a model that was three weeks old. China's Z.ai released GLM-5.3, now tied for strongest open model on independent benchmarks and second only to Claude Opus 5 on agentic tasks, at 68 cents per task. Why it matters: most business workloads (summarising, drafting, classifying, extracting) do not need a frontier model. Moving them down a tier is now quality-neutral and often a tenfold cost saving. (deepmind.google)

The European angle

Mistral shipped the most practically useful European release of the fortnight: Agentic Search, a retrieval layer that lets a model search, open and navigate documents in steps instead of grabbing everything in one pass. On the FinanceBench test its accuracy went from 26.7 to 86 percent, with lower latency and a third fewer tokens. If you have tried a document assistant on your own contracts, manuals or tender files and found it disappointing, this is the technique that fixes it, and it comes out of Paris. Closer to home, Ghent-based Dropsolid won a VLAIO innovation grant to build sovereign AI agents. (mistral.ai, itdaily.be)

Also this week, in one line each

• SpaceX closed its 60 billion dollar acquisition of Cursor, the largest startup exit on record. (techcrunch.com)
• Stripe is buying model router OpenRouter for over 7 billion, Nvidia is taking Poolside's model factory, and Hugging Face is reportedly in talks at 13 billion. (techcrunch.com)
• Anthropic's own August risk report disclosed that one of its safety filters was inactive for eleven months, a reminder to verify vendor safety claims rather than assume them. (thenextweb.com)
• Microsoft added Teams policy controls for meeting recording and notetaking bots, so you decide which AI notetakers may join your calls. (office365itpros.com)
• A Stanford study found AI is hitting entry-level jobs hardest, which makes how you train juniors a strategy question. (arstechnica.com)

Try this

Claude Code has a new /design command that produces UI mockups directly in the terminal. Next time you scope a small internal tool, run it live in the meeting: describe the screen, generate a few variants, and let the people who will use it point at the one they want. It turns a two-week mockup round into twenty minutes.

---

Researched and drafted with AI, then fact-checked, edited and approved by Edward Berger, who holds editorial responsibility for this edition.

Want to know where AI makes YOUR business money? Our AI Opportunity Scan shows you in one afternoon of your time: a concrete roadmap, a working demo on your own process, and a fixed-price proposal. One fixed fee, fully credited against the build. https://www.madisonunderwood.com/ai-opportunity-scan