I have a question for you. When was the last time you actually read the price page of the AI tool you are using right now? Not the landing page, the one with the friendly gradient and the cartoon robot. The other one. The fine-print one that lists what happens after the credits run out.
If you cannot answer that question from memory, I want you to keep reading, because the people selling you subsidized AI tokens right now are running a playbook that works every single time, and most of the folks using those tools are going to feel it in their budget before the year is out.
The subsidized token discount always ends
Here is the playbook. A lab launches a frontier model. They attach a fat credit giveaway, sometimes a whole year of “free” usage, sometimes a steep flat monthly rate that has no business being that cheap. You sign up because, well, it is cheap. You wire it into your daily work. Maybe you even told your team how smart you were for catching the wave. Then the credit pool shrinks quietly. The model gets slightly dumber on the tasks that matter to you, but only by enough that you do not name it as a regression. Your usage starts burning through whatever allowance is left faster than it used to. Eventually the price goes up, or the offer just ends, and by then your workflow is married to the tool.
That is a rug pull. No one calls it that in the keynote, but it is.
The reason it works is the discount phase trains you to use the model for everything. Code reviews. Email drafts. Customer support summaries. Casual questions to kill time between meetings. Once the meter starts running for real, all of those habits cost somewhere between three and ten times more, and you do not get the discount brain back. You only get the bill.
A local model covers most of what you actually do
Here is the part where I get to be the bearer of useful news. You do not need a frontier model for most of what you are sending to a paid API today. A lot of that traffic is summarization, formatting, classification, short rewrites, and boilerplate drafting. A 7B or 14B parameter model (a model with roughly 7 to 14 billion internal settings that determine its behavior) running on a Mac you already own handles that comfortably, and you already paid for the Mac.
The part nobody tells you is that the wrapper matters more than the model size. The harness that routes requests, formats prompts, remembers your preferences, and decides when to call out to a bigger model is where the real work happens. Two setups with the same base model can feel like night and day because of how the wrapper is wired together. If you are going to invest in local AI, invest in the harness, not the parameter count.
Once you have a small local stack tuned to your work, you stop refreshing the pricing page of your favorite lab every other Tuesday. You stop caring about which company just downgraded which model in a release note. Your bill becomes your electricity bill, plus a one-time hardware bill if you do not already own a machine that can run a 14B model at a useful speed.
Inboxes are the new chat windows for agents
Two big companies shipped an interesting coincidence this month, and coincidences in this space are usually signals. Both of them independently shipped the same idea, which is an email inbox built for long-running agents that work on your behalf, not just chat windows you poke at.
Cloudflare released an agentic email client that runs on its Workers platform. Incoming mail routes through their Email Routing product, each mailbox lives in its own Durable Object (a small persistent compute unit with its own dedicated database) backed by SQLite, and attachments drop into R2 storage. The agent scans your inbox, looks up threads, and composes replies, but does not send anything on its own. You approve every outbound message before it leaves.
AWS shipped a similar project under Apache 2.0 (a permissive open-source license that lets anyone use, modify, or resell the code) that is explicitly local-first. Finished work shows up as unread threads in your inbox. Anything that needs your decision gets flagged. The whole thing runs on your own hardware through LangGraph (an open-source framework for building multi-step AI workflows) and DeepAgents.
Things to notice about both of these at the same time:
- Agents want persistent storage rather than a transient chat log, because conversations die when you close the tab and mail sticks around
- Human-in-the-loop approval before any send action is showing up as the default safety posture, not a buried settings toggle
- Self-hosted and local-first are starting to ship from the biggest vendors, not just the indie hacker crowd anymore
- Voice and email are both chipping away at the chat-only interaction model that has dominated the past two years
Chat served as a stopgap. The next interface of choice will be an inbox, because that is how humans already organize work that takes more than thirty seconds.
What I would tell past me about all of this
If I could send a message back to myself twelve months ago, it would be a short one. Stop designing your workflow around whatever credit allowance is loudest this quarter. Build the local stack first, treat the cloud API as a fallback for the things the local model cannot do yet, and stop memorizing what each lab is calling their pricing tier. None of that is the actual product. The actual product is whether the work still gets done when the discount disappears.
Two more things worth sitting with. First, when two large, unrelated companies ship the same pattern in the same month, that pattern is going to be everywhere in twelve months. The agent inbox is not a curiosity. It is the next surface area. Second, if you are running a team and your monthly AI bill has gone up more than the headcount on your team, the discount phase did its job and is now over. Time to decide what portion of that work is going back to your own hardware.
Trade-offs
Here is the honest take.
For one, local models are slower than frontier cloud models for the hardest prompts, especially long-context reasoning over a hundred pages. A 14B model on a Mac mini will not give you frontier-class coding ability, and it will not catch you up on a difficult math problem the way GPT-class models often do. You are trading top-end capability for predictable cost.
That all said, the agent inboxes have a real downside too. You are giving an AI the ability to act on your behalf over email, which means a hallucinated action (a confident but incorrect command the model invents) could send a reply to the wrong person or take a workflow step you did not approve. The approval gate helps, but only if you actually look at the draft and not just thumb the approve button out of habit. Treat it like handing your laptop to a friend and saying “fix the Wi-Fi” and you are about right on the trust model.
Last point worth making. The privacy story on local-first is good, but it is not automatic. If your local agent syncs to a cloud storage backend because that is the path of least resistance, you have not gained what you thought you gained. Read the data path before you trust the marketing.
What this looks like in six months
The shape of what is coming is reasonably obvious from here. The subsidized token tier is going to keep shrinking. Local model quality is going to keep climbing. The agent inbox pattern is going to spread to the major productivity suites within the year, whether or not they ship it themselves or buy a startup that does. The teams that come out ahead are the ones that already had their own harness running when the bills started arriving, not the ones who start scrambling in Q2 of next year when finance asks why the AI line item grew 400%.
If you only take one thing from this, take the harness seriously. The model is the part everyone has opinions about. The wrapper is the part that decides whether the work actually gets done.
The cheap tokens are not the product. The product is the wrapper you wrote.