The Forge — Armory Forge Systems — Signet Article #040
Ask yourself what it would be worth to have an assistant that adds an appointment to your calendar, reschedules a call when something comes up, and emails the other person to confirm.
Now ask what that's worth per month. Hold that number.
Because the two biggest AI companies in the world just answered it very differently — six days apart.
On September 3rd, OpenAI opened up GPT-6 Astra, calling it the most capable model it has ever broadly deployed, with headlines about the arrival of AGI. On the pricing page, Astra costs $10 per million input tokens and $50 per million output tokens — roughly two and a half times the price of the model it replaced.
On September 8th, Meta shipped Muse: a personal agent that reads your email, books your travel, fills out forms, negotiates a bill, turns a saved recipe video into a grocery list, and keeps working after you close the app. It's free to start, with paid tiers reported around $20 and $100 a month. The model underneath it, Muse Spark 1.3, lists at $1.25 in and $4.25 out — about a twelfth of Astra's output price.
And here's the part nobody at either company will say out loud: Meta is undercutting OpenAI with the weaker model, and betting that for most of what people actually ask an agent to do, weaker is fine.
The calendar test
Run through what an agent is actually asked to do in a normal week. Add a dentist appointment. Move Thursday's call to Friday. Book the flight. Cancel the subscription nobody uses. Fill out the form. Send the follow-up. Turn a recipe into a shopping list.
Not one of those tasks is a reasoning showcase. There's no decade-old math problem hiding in your calendar. The intelligence required is high enough — not highest available. And "highest available" has a price, which means it has a limit: the moment your agent gets expensive enough per move, you start rationing it. You use it for the big stuff and do the small stuff yourself.
A rationed agent is just a worse employee.
Meta understood something OpenAI's framing obscures. In chat, you pay per question. In agent work, you pay per loop — check, act, verify, retry, report. The same task can burn dozens of passes. That's not a rounding error; it's the entire economics of whether an agent is worth leaving on all day. Price isn't a detail of agentic AI. Price is a feature.
The catch nobody puts in the headline
There's a reason Meta's prices look impossibly low: they are, and the discount comes with a receipt. Meta's contributor tier runs about a 95% discount in exchange for the right to train on your prompts and outputs. That's a legitimate trade for a hobbyist. It's a non-starter for a business holding client records, payroll, or patient data.
Which is the first hint that consumer AI and business AI are different products wearing the same costume. Cheap intelligence is real. Intelligence you can actually put your company's data into costs more — not because the labs are greedy, but because the terms change the moment the data matters.
Intelligence is a dial, not a badge
Here's the correction the price war is forcing on everyone: model choice is not a loyalty, it's a routing decision.
You do not use one brain for everything. You use the cheap one for the ninety percent that's routine, a mid-tier one for the work that needs judgment, and the expensive one for the small slice that's genuinely hard. Paying frontier prices to add a calendar entry is like hiring a heart surgeon to change a lightbulb — technically capable, economically absurd.
The companies that figure this out will run agents that cost a fraction of what their competitors pay, doing the same work, because they stopped buying a superintelligence for every task and started buying the right intelligence for each one.
What Meta actually proved
Meta didn't just prove a cheaper model can be good enough. Meta proved the model isn't the product.
Look at what Meta built around that mid-priced model: every user gets their own dedicated virtual machine, where the agent, the data, and the account credentials live. A second agent sits on the boundary and approves everything that tries to leave. It keeps a memory of your preferences. It asks before it spends. The agent runs somewhere you don't administer, doing things you approve, with a gate you don't control.
That's not a chatbot with a subscription. That's a harness. And Meta's own announcement says it was modeled on the open-source agent OpenClaw — meaning the reference architecture for this category is now a public consensus, not a secret.
That's genuinely good news, and it should be read as a warning. The model is becoming a commodity with a falling price. The harness — the worker's identity, its permissions, its memory, the gate on what it can touch — is where the value moved. If your entire plan was "we have a better model," you don't have a plan.
Your business needs the best of both worlds
Everything above is about consumers. Meta's Muse is built for people: personal errands, personal accounts, one human. It has no concept of a department, a role, a permission boundary, a compliance requirement, or a company's books. It can't be your IT department. It can't answer your phones. It can't know your clients.
But the trust model Meta just normalized is the right one — your assistant's data and access sealed off from everyone else's, with the gate in the architecture rather than in a policy document. That part deserves to be the industry standard. It should have been from the start.
What's missing is the business half: workers that hold a role, know the company they work for, and can be trusted with real records — and intelligence economics that let a small business run a whole workforce of them instead of one premium assistant it's afraid to use.
That's what we're building, and it's the version we think this market actually needs: as secure as what Meta just handed two billion people, and considerably more efficient for a business.
We'll have more to say about it soon.
The price war between Meta and OpenAI is one of the best things to happen to small business AI. You don't have to pick a side. You get to pick the right intelligence for each job — and for the first time, that's a choice you can afford.
Somebody still has to assemble it, wire it into your phone line and your books and your inbox, and make sure it's yours.
That's not a consumer app. That's a company.
Armory Forge Systems builds AI workers that run real departments for small businesses. If you want the best of both worlds — not the most expensive one — let's talk.
