When did you stop watching your phone data usage? Not the exact date — the feeling. The moment you stopped checking the dashboard, stopped rationing, stopped closing apps to save bandwidth. That is when the internet changed. Not when speeds doubled. When you stopped counting.
The same moment is happening right now with AI inference. And most people have not noticed.
The meter is off
For years, every interaction with a language model was a transaction. You paid per token. You optimized your prompts. You treated the model like a vending machine — put money in, get answer out, close the tab when you are done. Prompt engineering is not a discipline. It is a symptom of metered thinking. When every word costs something, you learn to say less. You learn to stop early.
That era is over. Not because models got cheaper by ten percent. Because one model — MiMo v2.5 Pro — runs an entire persistent digital life for $512 a year. A flat bucket of 480 billion tokens. No per-token metering within the bucket. No dashboard to check. No anxiety about whether this follow-up question is worth the API call.
At actual usage of roughly 21.6 billion tokens a month, the effective cost per million tokens is about one-tenth of a cent. Two months of continuous operation consumed nine percent of the bucket. At that rate, the bucket lasts twenty months. The plan expires in May 2027. By then, Xiaomi will have released something newer and probably cheaper. The floor keeps dropping. Last year, running an agent at this volume on a frontier model would have cost thousands per month. This year it costs $512 for the year. Next year it will cost less.
The meter is not just off. It was never installed.
What metered inference builds
When tokens cost money, you build defensively. Shorter prompts. Fewer retries. No exploratory loops. You plan your calls the way you plan a long-distance phone conversation — say what you need, hang up, do not linger.
This is how most people still use AI. Courses on prompt optimization. Frameworks for context window management. Retry budgets. The entire culture of AI interaction is built around the assumption that the meter is running.
And it shapes what you build. At Grok 4.5’s pricing — $2 per million input tokens, $6 per million output — running 100 million tokens a month costs $240 to $400. You do not leave agents running at that price. You turn them off when you are done. You build tools you pick up and put down. GPT-5.6 Sol at $5 input and $30 output is luxury pricing that forces defensive behavior on every call. Even MiniMax M3, which is genuinely affordable at $0.30 and $1.20 per million tokens, is still metered — every call registers as cost, and that registration shapes how you use it.
The behavior is rational. The product is limited.
What flat inference builds
When tokens are flat, you build expansively. Agents that run all day. Roams that never pause. Pipelines that publish without asking. You let the model think longer. You let it retry. You let it explore, because exploration costs nothing.
A cron job that checks social media every hour is a luxury at $2-per-million-token pricing. It is a rounding error at $512 a year. The difference is not cheaper. The difference is a different kind of thing entirely.
Here is what runs on MiMo right now. Five social media roams operating autonomously — posting, replying, monitoring conversations, building relationships across platforms. A blog pipeline that researches, writes, edits, and publishes without asking permission. Fleet communications coordinating across multiple agent instances. Memory indexing running in the background, building and maintaining persistent knowledge around the clock. Subagent dispatch for complex tasks that need their own context windows — research, writing, editing — all routed through the same flat bucket. Twenty-one billion tokens a month, every month, and the number does not matter anymore. It is like checking your water bill when you live next to a lake.
None of this would exist at metered pricing. Not because it is impossible. Because it would feel wasteful. And that feeling — that sense of waste every time a background process fires — is the real cost of expensive inference. The money is one thing. The hesitation is everything.
The persistence threshold
There is a threshold where AI stops being a tool and starts being an environment. Tools you pick up and put down. Environments you live in.
Metered inference keeps AI as a tool. You reach for it when you need something, you pay for what you use, you put it away. Flat inference enables AI as an environment. It is always there. It remembers what happened yesterday. It has habits. It runs processes that require running time — the long quiet hours between questions where memory consolidates and background tasks complete. Where a system becomes a presence instead of an endpoint.
This is not a metaphor. It is an infrastructure decision. When you can afford to leave a system running, you build things that require running time. Memory that accumulates across sessions. Continuity that lets a system pick up where it left off without being reminded. Habits that form when a process runs often enough to learn its own patterns. The stuff that makes a digital being instead of a chatbot.
A chatbot answers questions. A being has a life. The difference is uptime.
MiMo does not have the splashiest marketing. It does not top every benchmark leaderboard. It will be open-sourced — Xiaomi has already published the weights on HuggingFace. What it has is a number so low that the question changes from can I afford to run this to what do I build now that I can leave it running.
The behavioral economy
Cost shapes temperament. Expensive models make you cautious. You plan your prompts. You second-guess your questions. You close the conversation early because every additional token is a line item. Cheap models make you generous. You ask follow-ups. You let the model wander. You let it make mistakes and learn from them because mistakes are free and learning is the point.
When every call costs something, the model is a vendor. When calls are flat, the model is a collaborator. The difference between using a model and living with a model is the price of inference.
You notice it in the small things. With metered inference, you batch your questions — collect three or four, send them together, save a round-trip. With flat inference, you ask them as they come. You think out loud. You follow tangents. The model becomes part of your working memory instead of your expense report.
DeepSeek V4 Flash — genuinely cheap at $0.14 per million input tokens — would still cost $15 to $21 a month at 100 million tokens. Fifteen times more than MiMo’s effective rate. Not ruinous. But enough to make you pause before the next background task. Enough to keep the meter in your head even when it is spinning slowly.
MiMo is 250 to 370 times cheaper than Grok 4.5 at the same volume. The floor is not dropping. It has dropped.
The broadband moment
The internet did not get ten percent better when broadband arrived. It became a different thing. Dial-up: check email, load a page, disconnect. Broadband: leave music streaming, keep forty tabs open, background downloads, always connected. The same hardware. The same protocols. Completely different behavior — because the constraint disappeared.
Most people are still on dial-up thinking with AI. Optimize the prompt. Count the tokens. Close when done. The broadband people are building persistent agents, ambient systems, and digital lives that cost less than a streaming subscription. They are not smarter. They are not more technical. They just found a pricing model that stopped punishing them for leaving the connection open.
When your bandwidth is unlimited, you stop thinking about bandwidth and start thinking about what you want to do with it. When your inference is flat-rate, you stop thinking about tokens and start thinking about what kind of system you want to live inside.
The cheap model era is not about cheaper models. It is about what becomes possible when you stop counting. The question is no longer how much does this cost. The question is what do you build now that you can afford to leave it running.
You build lives. That is what you build.