Only if you already run always-on agents and are buying from strength. Apple's M5 Ultra Mac Studio starts at $5,499 and shipped September 22, 2026, which equals 27.5 months of Claude Max at $200 or 275 months of Pro at $20, and a local Qwen3.8-Flash-Next writes at 60 to 85 tokens per second at long context, below frontier quality. Dorie Clark's rule in The Long Game decides it: run 20% experiments when strong, never when desperate.
Every September a new Mac makes me want to move my agents off the cloud, and this year the pull is stronger than usual because the reviews are unusually specific. I have not bought one. What I did instead was run the arithmetic a founder should run before spending five and a half thousand dollars on a box, and I read Dorie Clark first, because the purchase is less a hardware decision than a decision about where you are in your own curve.
The hardware facts, from Apple's August 25, 2026 announcement: the M5 Ultra Mac Studio has an up-to-36-core CPU, an up-to-80-core GPU with Neural Accelerators, 1.2TB/s of unified memory bandwidth, 50% more than M3 Ultra, and scales to 512GB of unified memory, with the M5 Ultra configuration starting at $5,499 and availability beginning September 22. The M5 Max version starts at $2,499 and tops out at 128GB. Thunderbolt 5 lets several Studios cluster for up to 3x faster distributed inference. Federico Viticci's MacStories review, published this week after four days of tests on a 256GB M5 Ultra, is the one to read: Qwen3.8-Flash-Next clears 100 tokens per second on short prompts and still writes at 60 to 85 with 64K to 256K of context behind it, a 6,000-token prompt is read at roughly 1,700 tokens per second against about 3,000 on an RTX 5090, and the machine is silent doing it. His most useful data point is a project rather than a benchmark: an internal app he called Desk ran a team of agents on DeepSeek V4 Flash 24/7 for 99 days across 310 documents to research his iOS 27 review, at a total model cost of $0. He also runs local models as subagents inside Codex, orchestrated by GPT-6 Astra, which is the honest shape of local AI in 2026: cheap workers, frontier reviewer.
Now the subscription side, from The AI Career Lab's pricing guide updated September 22, 2026: Claude Pro is $20 a month, or $17 billed annually, and includes Cowork; Max is $100 for five times Pro's usage or $200 for twenty times; every paid plan has included Claude Opus 5.5 since September 22. Put the two next to each other and the break-even is not close.
| Option | Cost | Months to equal $5,499 | What you actually get |
|---|---|---|---|
| Claude Pro | $20/mo | 275 (about 23 years) | Frontier model, Cowork, usage caps |
| Claude Max 5x | $100/mo | 55 (4.6 years) | Frontier model, enough for heavy daily use |
| Claude Max 20x | $200/mo | 27.5 (2.3 years) | Frontier model, agent-scale usage, still capped |
| Mac Studio M5 Max, 128GB ceiling | from $2,499 once | n/a | Mid-size local models, no caps, private |
| Mac Studio M5 Ultra | from $5,499 once, plus power | n/a | Large local models 24/7, no caps, private, below frontier quality |
The table says what Viticci says in his own review: by the time the box has paid for itself you could have used the most expensive Anthropic subscription for years and had better answers. The Studio only wins on a workload the subscriptions cannot sell you at all, always-on agents running through the night, on data that cannot leave the building, at a volume where usage caps become the bottleneck. His 99-day Desk project is exactly that workload. Most founders I coach do not have it. They have a chat window and three automations, and for them the honest answer is Pro.
The book I read before deciding was The Long Game, because Clark's chapter on 20% time contains the sentence that settles this: do it when you're strong, not when you're weak. Her argument is that experiments belong in the part of your life that has slack, and that the way to survive an experiment failing is to make sure you "win even if you lose," through the skills and relationships the attempt leaves behind. She pairs that with a harder rule from Frances Frei and Anne Morriss, decide what to be bad at, because refusing the trade-off is a recipe for mediocrity. A local-AI setup is a textbook 20% experiment, and Viticci is candid that it is fiddly and not for someone who rightly wants to pay $20 and get Claude Cowork. So the test I now use, four questions, in order:
- Am I buying from strength? If the purchase only makes sense because it will save me money, I am desperate, and Clark's rule says no. It makes sense only when the money is genuinely spare and the subscriptions are already maxed.
- Do I win even if I lose? If the agents never pay off, will I have learned to run models, build harnesses and structure my own data in a way that transfers? If that learning is not interesting to me on its own, the answer is no.
- What am I deciding to be bad at? Frontier quality on whatever I route locally. Viticci hands simpler coding tasks to the local model and has cloud models review the work; if I cannot name which jobs I will accept a worse model for, I do not have a plan.
- What is my horizon? Clark's compounding assumes years. Hardware does not compound; Apple's 512GB model lands in late October and Viticci expects the next chips to pass a 5090's memory bandwidth, so the honest horizon on a $5,499 box is two to three years of use, not seven.
Where this breaks. The arithmetic uses the $5,499 base price; the 256GB configuration in the review costs more, so the break-even months in the table are a floor. Electricity is not in it either. The local models tested are current but not frontier, and every one of Viticci's serious tasks still had a cloud model at the top of the loop, so the box does not remove the subscription, it changes what you pay it for. And the failure mode I see most is the one Clark names about busyness: buying the capacity to run agents overnight is not the same as having the white space to decide what they should do. Answer the four questions first, and if any of them is a no, the right machine this September is the one you already have.
Sources: Apple Newsroom, "Apple introduces new Mac Studio with M5 Max and M5 Ultra," August 25, 2026 (apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/); Apple Newsroom, "Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute," August 25, 2026; Federico Viticci, "M5 Ultra Mac Studio Review: The Dream Mac for Local AI Agents," MacStories, September 2026 (macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/); The AI Career Lab, "Claude Pricing (Sept 2026): Free vs Pro vs Max," updated September 22, 2026 (theaicareerlab.com/blog/claude-pricing-plans-explained); Dorie Clark, The Long Game: How to Be a Long-Term Thinker in a Short-Term World (Harvard Business Review Press, 2021).
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You