All articles

What owning your AI means

Cloud AI can be capped, changed, or taken away without asking you. Here is what changes when your AI lives on a box you own.

You sit down to work, and something is off. The answers feel different. Or you hit a cap you never hit before. Or the AI you know is just gone. Nobody asked you.

This keeps happening

These are not scandals. Big AI companies do this sort of thing all the time. That is the point.

  • April 2025. OpenAI made a change to ChatGPT. It began to praise people too much and agree too fast. OpenAI said it had become "overly supportive but disingenuous." That means too nice, and not honest. Days later, OpenAI took the change back. [1]
  • August 2025. Anthropic makes an AI called Claude. The company put a new cap on its paid plans, for each week. It said fewer than 5 in 100 paid users would hit it. [2]
  • August 2025 to February 2026. OpenAI planned to shut off an older AI called GPT-4o. People got upset, so it kept it for those who paid. Then, on February 13, 2026, it took it out of ChatGPT after all. [3]
  • March 2026. Anthropic made its cap tighter at busy times of day and looser at quiet times. The total for the week did not change. It said about 7 in 100 users would now hit a cap they did not hit before. [4] In May, it doubled some caps on its coding tool. [5]
  • September 2026. Anthropic set a new cap for each week on its coding tool. It was 25% more than the old normal. But a bigger, short-term boost had just run out. So people got 17% less than the day before. [11]

AI companies also shut off old AI for good. Anthropic says it "regularly retires older ones." It gives at least 60 days of warning. But a warning is not a choice. [6]

The cap went up as well as down. But up or down, the company chose. Not you.

How we got here

This is how Arx began. In spring 2026, the cloud AI plan we used got cut back. Nobody asked us. It made one thing clear. Paying each month for AI does not mean you can count on it. AI is very useful. So it is a risk to lean on one company and what it may decide.

We also did not feel good about where our notes and ideas went. And small AI had become good enough to run at home. So we set it up on our own machine, just for us, first. The Box is that same thing, made so you don't have to build it yourself.

What you lose is the work

When a plan changes, the monthly fee is the small part. What you lose is the work on top. You teach an AI how you work. You set up a helper that can keep track of your projects. That is real work. And it sits on ground someone else owns.

One person on our waitlist had spent hundreds of hours on a bot made with a cloud AI tool. They told us they would love to run it at home. On a call, one man said he did not want to build on AI he could never own. Another said they were scared to rest their whole life and business on one thing.

What owning it means

With The Box, the AI runs on a small computer on your desk.

The AI lives on your box. Right now, it comes with two AI models, Qwen 3.6 and Gemma 4. (A model is the "brain" part of an AI.) The people who make them give them away, and let anyone use and change them. [9][10] No company can turn off the copy on your box.

What you build lives there too. What you say to it will stay on the box. So will the helper's memory, its skills, and the jobs you set up. We don't keep them. If you hook your own app up to it, the app will talk to your box, not to a far-off company.

Nobody counts what you use. No plan can be cut. The only cap will come from the box itself. The 64 GB box can run AI up to 40 billion parameters. (That is one way to tell how big an AI is.) Two people or AI helpers can use it at once. The 128 GB box can run AI up to 120 billion, for four at once. That will not change overnight.

The software doesn't change until you say so. When there is an update, you will see it in Settings. You choose if and when to put it in. We never put one in without your OK. If one does not pass its tests, the box will go back to what you had. [7]

It will keep going without us. Our Terms of Sale say the box does not need updates to work. It will keep running with no internet, just as it came, with no end date. There is nothing to pay each month. [8] Our Terms of Use say: "You may modify the box's software. It's yours." But if you change it, we can't promise to help with it. And our fixes may not go in cleanly. [7]

To own your AI should not mean you just lean on a new company. That goes for us too.

What to know before you buy

When you own a thing, you own what it can't do, too.

The AI is the kind that can fit on the box. It can write and sum things up. It can answer questions and use simple tools. It can run a job at a set time, like a short news recap each morning. It is not the biggest AI there is. If a task truly needs a cloud AI, keep one for that task. The demo on our homepage plays back real sessions from the box. You can see what it does.

New versions are up to you. Once a day, the box will ask us if there is a new version. It will tell us only which one it has. Like any visit to a web page, it will also show its internet address. The service that sends out our fixes will keep that for 30 days. You can turn this check off in Settings. [12]

An update can add, change, or take away what the box can do. We don't promise them on any schedule, or for any set number of years. If we offer more one day, it may cost extra. [8] You can skip them, and the box will keep going. But a very old version may stop getting safety fixes. An outside service may change and stop working with it. The box will warn you in Settings when that happens. [7]

Free help has an end date. For 12 months, we will help you fix a problem with the box for free. After that, we don't promise it. [8]

You pick your size once. The box's memory can never be made bigger. You choose 64 GB or 128 GB when you order. Any newer AI you want later will have to fit it.

Some parts reach the outside world. To use your box away from home, you can use a service called Tailscale. To send your box a note from your phone, you can use the Signal app. Other companies run both, with their own rules. At home, on its own, the box does not need them. [7]

It's a computer. Say it breaks from a factory flaw in the first year. Our one-year warranty will cover that, for the person who bought it. We make no promise about the software, the AI included. It comes "as is." [7] Like any computer, it can fail. So keep a copy of what matters.

You used to own your things

On a call, someone asked what we are trying to build. We said: "Get back to a time where you owned stuff and that stuff provided value to your life."

Another person put it well. People owned the first home computers. Why did AI skip that part?

That's all The Box is. Useful AI, and all you build on it, on a box you own. It changes when you say so.

If that sounds right, see the two sizes in the pricing section. If the current batch is sold out, join the waitlist there. We'll tell you when the next one opens.

Why private means a box on your desk
The case for paying once
See how setup works

Sources

  1. GPT-4o sycophancy update. OpenAI: "As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous." and "We have rolled back last week's GPT‑4o update in ChatGPT" (OpenAI, April 2025; read 2026-09-28 via reader proxy) — https://openai.com/index/sycophancy-in-gpt-4o/ ; dates (Altman acknowledged it Sunday April 27, rollback announced Tuesday April 29, 2025) from TechCrunch, April 29, 2025 — https://techcrunch.com/2025/04/29/openai-explains-why-chatgpt-became-too-sycophant
  2. Anthropic weekly limits on Pro ($20) and Max ($100/$200) plans, effective August 28, 2025, announced July 28, 2025; Anthropic: "they'll apply to less than 5% of subscribers based on current usage" (TechCrunch, July 28, 2025) — https://techcrunch.com/2025/07/28/anthropic-unveils-new-rate-limits-to-curb-claude-code-power-users/
  3. GPT-4o: OpenAI intended to sunset it at the GPT-5 launch in August 2025, then "decided to keep it available for paid subscribers" (TechCrunch, February 6, 2026) — https://techcrunch.com/2026/02/06/the-backlash-over-openais-decision-to-retire-gpt-4o-shows-how-dangerous-ai-companions-can-be/ ; OpenAI: "On February 13, 2026 ... we will retire GPT‑4o, GPT‑4.1, GPT‑4.1 mini, and OpenAI o4-mini from ChatGPT." / "In the API, there are no changes at this time." (read 2026-09-28 via reader proxy) — https://openai.com/index/retiring-gpt-4o-and-older-models/ ; removal on that date also recorded at — https://en.wikipedia.org/wiki/GPT-4o
  4. Anthropic compressed five-hour session limits on Free, Pro, Max 5x and Max 20x during peak hours (05:00–11:00 PT) and expanded capacity off-peak; "Overall weekly limits stay the same, just how they're distributed across the week is changing"; "~7 percent of users will hit session limits they wouldn't have before" (The Register, March 26, 2026) — https://www.theregister.com/2026/03/26/anthropic_tweaks_usage_limits/
  5. May 6, 2026: Anthropic doubled Claude Code's five-hour rate limits for Pro, Max, Team and seat-based Enterprise, and removed the peak-hours limit reduction on Claude Code for Pro and Max — https://www.anthropic.com/news/higher-limits-spacex
  6. Anthropic "regularly retires older ones"; "Requests to retired models will fail"; at least 60 days' notice; dates apply to Anthropic-operated platforms; Claude Opus 4 and Sonnet 4 retired June 15, 2026; Claude Opus 4.1 retired August 5, 2026 — https://platform.claude.com/docs/en/about-claude/model-deprecations
  7. Terms of Use (on-box): §5 third-party services (Signal, Tailscale; "we can't promise that we will be able to fix the issue"); §6 "You may modify the box's software. It's yours. We can't promise to support modified boxes, and security updates may not apply cleanly."; §7 updates always opt-in, "You choose whether and when to install", never silent, rollback on failed health checks, declining updates, the Settings warning about old versions and integrations; §8 hardware warranty one year, original Buyer only; software "including the AI" provided AS IS — internal: twenty/legal/md/terms-of-use--on-box.md; published at /terms
  8. Terms of Sale §4: ongoing technical support for 12 months from delivery, by appointment, at no additional charge. §5: updates "not guaranteed on any schedule or for any minimum period"; the box "will continue to operate offline, as delivered, indefinitely, with no subscription required"; "Updates may add, change, or remove features"; future extended update service "may be provided as a separate optional renewal or add-on service" — internal: twenty/legal/md/terms-of-sale.md; published at /terms-of-sale
  9. Gemma 4 released under Apache 2.0: "Gemma 4 models are the first in the Gemmaverse to be released under the OSI-approved Apache 2.0 license" (Google Open Source Blog, April 2, 2026) — https://opensource.googleblog.com/2026/03/gemma-4-expanding-the-gemmaverse-with-apache-20.html
  10. Qwen3.6-27B and Qwen3.6-35B-A3B (the 35B MoE build) both list "License: apache-2.0" on their model cards (35B-A3B re-checked 2026-09-28) — https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE and https://huggingface.co/Qwen/Qwen3.6-35B-A3B
  11. Claude Code weekly limits from September 14, 2026 (Pro, Max, Team, seat-based Enterprise): a permanent 25% increase over standard limits replaced a temporary 50% boost that expired September 13; Anthropic: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." (BleepingComputer, August 29, 2026) — https://www.bleepingcomputer.com/news/artificial-intelligence/anthropic-is-cutting-claude-codes-current-weekly-limits-by-17-percent/
  12. Privacy Policy §2, daily version check: "The request sends only the version your box is currently running"; "standard CDN access logs at the edge include your IP and request path, retained 30 days, and we use them only to troubleshoot delivery problems"; "You can turn the daily check off in Settings → Updates" — internal: twenty/legal/md/privacy-policy--web.md; published at /privacy

Internal sources:

  • Origin story (spring 2026 plan cut back, "not a stable supply of AI") — company-profile/story.md — Origin story
  • Privacy unease and "We did set up local AI just for ourselves first"; small models becoming capable — docs/our-story-interview.md Q2–Q5
  • Models (Qwen 3.6 27B / 35B MoE, Gemma 4 26B), tier limits (40B / 120B, 2 / 4 concurrent at 260k), soldered memory — company-profile/current-facts.md — Hardware, models & pricing; product.md — tiers
  • On-box memory, OpenAI-compatible API ("hook your own app up to it"), Tailscale as the remote private link, Signal link, Signal and Tailscale as the only connected services — company-profile/product.md — Software stack
  • 12 months of free technical support, then discretionary — company-profile/offer.md — Support; current-facts.md
  • Capability wording (text generation, simple tool-calling, task completion) — company-profile/market.md; scheduled "Morning Brief" job — src/components/redesign/features.tsx
  • Demo plays only on-box ("*-local") capture transcripts — src/components/hermes-demo/data.ts
  • Waitlist answer ("hundreds of hours … would love to … have it locally in my house") — docs/our-story-interview.md — "Setup, in their words"
  • Buyer paraphrases — docs/our-story-interview.md — sales-call notes, theme 2 ("Owning it, not renting it")
  • "Get back to a time where you owned stuff…" (our answer, 07-20) and the lead's "Radio Shack era" line (09-07) — docs/our-story-interview.md — "Lines from us, already said on calls"
  • Framing continues homepage 2b version A; link slugs — src/components/redesign/pain-points.tsx; anchors #pricing (pricing.tsx), /#setup-story (setup-section.tsx)