Ask a cloud model one question and the token bill is small enough to ignore. Wire that same model into something that runs on its own — a workflow that fires on every chat message, checks a document on upload, or triages your inbox on a schedule — and you've turned an ignorable cost into a metered one that scales with how often you actually use the thing. The more useful the automation, the more it runs, the more it costs. That's backwards.

What "pay per token" actually means

Take OpenAI's own published pricing as of this week: the base GPT-5 API runs $1.25 per million input tokens and $10 per million output tokens. The flagship GPT-5.6 tier is higher still — $5 in, $30 out per million. Every request you send and every reply you get back is metered against those numbers, no exceptions.

That's fine for occasional use. It stops being fine the moment you automate something. A workflow that summarizes every document you upload, or replies to every message in a channel, isn't sending one request — it's sending one every single time it fires, indefinitely, for as long as you keep using it. The pricing model is built around usage staying occasional. Automation is the opposite of occasional by design.

What runs locally instead

Aegis loads models as plain GGUF files and runs inference through llama.cpp (via llama-cpp-python), directly on your own CPU or GPU — no request ever leaves your machine, and there's no API key to meter it against because there's no API call to make. You download a model once from the LLM panel, and every chat, every workflow run, every document it processes after that costs exactly what it already cost you to own the hardware. Not $0.01 more per run, not $10 more for a heavier one — nothing, because nothing is being billed.

This isn't a pricing trick or a free tier with a catch. It's a different shape entirely: the meter simply doesn't exist.

Where this actually matters: your daily workflows

This is the part that only shows up once something runs more than once:

A workflow triggered on every chat message — extract, search, reply, no per-message charge no matter how many messages come in that day.
Document Q&A on everything you upload — attach a report, a contract, a set of notes, and it's chunked, embedded, and searchable the moment you need it, run as many times as you want.
A reranking step on every search — a proper two-stage retrieval pipeline (search wide, rerank down) is usually left out of local tools specifically because it's an extra inference call, and extra inference calls cost money on a metered API. Locally, "extra" just means "a few more milliseconds."

None of these are exotic. They're the ordinary, repetitive parts of a workday — the parts worth automating in the first place, and the exact parts that get expensive fastest on a per-token bill.

Where we're still honest about the tradeoff

A model running on your own machine isn't the same model as the largest cloud ones. For the hardest reasoning tasks, a frontier API model will still outperform what fits in consumer RAM — that's real, and we're not going to pretend otherwise. You also need the hardware to hold the model (Aegis's Memory Hub shows exactly what a given model costs you in RAM before you commit to it), and the first download of a model takes real time and disk space, unlike an API call that's ready instantly.

What you get in exchange is a cost curve that stays flat no matter how hard you lean on it. For the daily, repetitive work — chat, document search, reranking, the stuff that actually benefits from running in a loop — that trade is usually the right one.