Local AI for small teams

A five-person studio adds a sixth person and the AI bill goes up again. That's how per-seat pricing works: every hire is another monthly line item, and the price can change without your say-so. Buying the tool once, per Mac, breaks that pattern. Outlier runs language models on each teammate's own machine, so it's a cost you pay once, not a seat that renews forever.

The per-seat problem

Cloud AI is priced like a chair at a desk. You rent one per person, every month, and the meter never stops. For a 2-to-15-person shop that math compounds fast: a handful of seats at $20 to $200 each grows with every hire and can be repriced at the vendor's discretion. Whole departments are feeling it. Axios called it "AI sticker shock" in May 2026, and a May 30 report described corporate America starting to ration AI as cost skyrockets. A May 17 piece put it plainly: AI subscriptions are "a ticking time bomb for enterprise."

The part that stings for a small team is that you're paying on terms you don't set. Cancel a seat and the work behind it is gone. Keep it and the bill rides up with payroll.

What changes when each Mac owns its AI

Outlier is a Mac-native app. One signed download, no account, no terminal, no Docker. The model runs on the machine in front of you, so the cost is tied to the machine, not to a renewing seat. Hand a teammate a Mac and the AI comes with it, no monthly per-head charge behind them. Three things shift for a small team:

• Flat, predictable cost. You buy it once per Mac. The free Nano and Lite tiers cost nothing and need no account; Pro is a one-time lifetime option from $99, or $20/mo if you'd rather. The number doesn't climb every time you hire.
• Data residency by default. Each person's prompts, code, and client files are processed on their own machine and stay on their disk. Nothing routes through a vendor, so there's no business-associate chain to manage and nothing landing in someone else's training set.
• No caps mid-deadline. Nobody is drawing from a shared, rationed pool. One person running a long job at 11pm doesn't burn the team's quota, because there isn't one. A real Hacker News thread is literally titled "Optimizing my sleep around Claude usage limits." That's the world you leave behind.

The math: five seats vs five Macs

Take a five-person team, illustratively. On a mid-tier cloud plan around $30 a month per seat, that's roughly $150 a month, or about $1,800 a year, and it renews every year, climbing as you add people or as the vendor adjusts pricing.

The same five people on Outlier can each start free on Nano and Lite. For the stronger models, lifetime Pro is a one-time $99 per person while the Founding 200 seats last, then $200. Five installs is a one-time cost in the few-hundred-dollar range, and next year it's $0. The cloud line item, by contrast, shows up again every January.

What stays on each machine

Because there's no API call, there's no third party in the loop. The prompt, the document, the response, the whole history all stay on that person's Mac. For a consultancy handling client material, that means no vendor terms of service to read on behalf of every client, and no central workspace pooling everyone's work into one breach surface.

Want proof? Any teammate can turn off wifi and keep working. Chat, document analysis, and project memory all run offline once the model has downloaded the first time. Outlier runs zero telemetry on inference and publishes its weights openly on HuggingFace, so this is checkable rather than promised.

The honest limits

This is not an enterprise IT product. There's no central admin console, no SSO, no MDM fleet management. It's a tool each teammate installs and owns, which is the right shape for a 2-to-15-person shop and the wrong shape for a 500-seat rollout. If you need centralized provisioning, this isn't it.

The hardest frontier tasks still favor the big cloud models. If one person does deep research or heavy reasoning all day, keep a single cloud seat for them. The argument here isn't "drop the cloud entirely." It's "stop renting five seats for work that runs fine on the five Macs you already own." On an M1 Ultra, Core 27B runs around 20.7 tokens per second where cloud flagships land near 80 to 100, so it's a usable drafting speed, not the quickest thing on the market.

For most of that daily work, the owned tool holds up. In a 54-prompt comparison, Outlier's local Core 27B matched Claude Opus on 98.9% of the rubric checks overall, and on all nine of the hardest tests: a chess engine, Raft and Paxos, zero-knowledge proofs. You can read the full benchmark and decide for the work your team actually does.

Frequently asked questions

Pricing figures for other AI tools are illustrative, used to show the per-seat math, and not a specific competitor quote. Confirm current prices with each vendor.