Three pricing pages in five weeks
On October 12, 2026, xAI moved GrokBot's agent functionality behind X Subscription Premium — $30 a month. The chatbot stayed free. The part that actually does work did not. Twenty-four hours later, Apple priced HomeView at $299 and attached a footnote to Apple Intelligence: server-side features carry daily usage limits, and expanded access "will be available for a fee in the future." Meta had already shown its arithmetic in September. Meta Muse: 100 million tokens a week free, $20 a month for 500 million, $100 a month for 3 billion. Three pricing pages in five weeks, one conclusion. There is no longer a free path to a functional home agent.
Context: what actually changed
For three years the pitch was uniform. A patient, always-on assistant in every living room, priced at zero because growth justified the subsidy. Venture capital paid for the inference, users paid nothing, and the platform bought market share with someone else's balance sheet. That is an acquisition curve, not a business.
What changed is not sentiment. What changed is that agents stopped being chatbots. A chatbot answers once and sleeps. An agent polls, retries, calls tools, and holds context across hours. It consumes compute while you are not looking at it, which means the vendor absorbs the variance. Flat free tiers cannot absorb variance. Metered billing exists because variance has to land on someone.
I have spent most of my career auditing meters. In 2018 I reviewed the 0x protocol v2 contracts line by line, examining how order routing priced execution. In 2020 I modeled emission schedules against locked value and watched the math resolve on schedule. In 2024 I reviewed custody architectures at asset managers and found key management concentrated in fewer hands than the marketing decks admitted. The lesson repeats. The instrument that measures the bill matters more than the bill.
The arithmetic behind the price sheet
Start with Meta, because Meta published numbers and numbers can be checked.
Meta Muse's paid tiers imply a unit price. $20 for 500 million tokens is $40 per billion, or $0.04 per million. $100 for 3 billion is $33 per billion. Prevailing API rates for mid-tier models sit one to two orders of magnitude above that. Either Meta Muse runs small models locally and charges a subscription for orchestration, or the paid tier is priced below marginal cost to anchor the platform. The free tier, at 100 million tokens a week, is roughly 433 million tokens a month — about $17.30 at Meta's own published rate. That is the subsidy per user, stated in the company's own arithmetic. The price sheet is a marketing document, not a cost model.
Apple's structure is harder to audit because Apple published no unit. HomeView costs $299. Apple Intelligence server-side features carry daily caps. Expanded access will cost something, at some point, set by Apple. There is no metering unit, no cap, no refund path, no notice period. What the buyer receives is a device plus an unpriced call option written against their future cash flow, exercisable by the seller on a date of the seller's choosing. Regulation by enforcement taught this industry that withholding clarity is cheaper than disclosing it, and the consumer version of that lesson is now shipping in a box. During the 2024 ETF custody review I learned to read footnotes before feature lists, because the footnote is where counterparty risk lives. Here the footnote is the product.
GrokBot is the most honest of the three and the most likely to be mispriced. $30 flat, no published token ceiling. Flat pricing on a variable-cost workload produces one of two outcomes: silent throttling, or a subsidy that gets repriced. Watch the latency. Watch the context window shrink quietly in a changelog. Flat fees do not stay flat when the underlying resource is auctioned.
The fourth layer of the compounding stack
That is the stack everyone now cites: agent runtime, cloud model API, device ecosystem. Each layer is sticky, each bills monthly, and consumers evaluate each in isolation. Add a layer nobody has priced. Agents that do work need to pay for things — tool calls, API credits, data, execution. That settlement increasingly lands on-chain, because on-chain is the only rail where a machine can hold an identity and a balance without a human signing a form. Every agent action that settles is a transaction: high frequency, low value, precisely the workload rollups were built for.
Which brings the bill forward. Blob space is a fixed resource auctioned every twelve seconds. Post-Dencun pricing made data availability cheap enough that agent traffic looked free. That window has a floor and a ceiling, and I have argued since the upgrade that saturation arrives inside two years. Follow the gas, not the narrative: agent metering is the demand curve that fills it. When blobs clear at the ceiling, per-call settlement costs double, and the compounding stack acquires a fifth layer no consumer has modeled — the cost of the ledger itself.
None of these platforms publish a usage attestation. Meta states token counts. Apple states nothing. xAI states a monthly figure and a gate. In 2021 I linked wallet clusters and demonstrated that roughly 40% of volume in the top NFT collections was wash trading by a small set of controlled addresses. I could prove it because the ledger was public and the counterparties could not edit it. Here, the token count exists only on the vendor's server, generated by the vendor's meter, displayed on the vendor's dashboard. You cannot audit a number that only one party can see. Code speaks louder than promises, and there is no code in a pricing page.
What the bulls got right
Here is the uncomfortable part.
The free tier was never a right. It was a customer acquisition cost, and it worked. Three years of subsidized inference bought these platforms a user base that would have taken a decade to assemble through retail channels. Ending it is not betrayal. It is the scheduled conclusion of a promotion.
More importantly, flat subscription may be the correct price for an agent, and metering may be the wrong abstraction. A per-token meter forces the user to ration cognition. You hesitate before asking your agent to retry, to poll, to hold a long context — and those are exactly the behaviors that make an agent worth owning. A meter converts a tool into a slot machine. A flat fee buys permission for the thing to think.
The blind spot in my own critique is that I am demanding a meter from vendors whose cost driver may be evaporating. NPUs in $299 hardware now carry enough throughput that on-device inference approaches zero marginal cost. If that holds, the rental thesis collapses for the cheap tier — and the platforms know it. That is why Apple gates the server-side features specifically. The local layer is the loss leader. The server layer is the business.
Publish the meter
The demand should be narrow and specific. Not "make it free." Publish the meter. A signed, timestamped, per-account usage attestation, exportable, is a solvable engineering problem and a trivial cost against the revenue it protects. The first vendor to ship one buys a trust premium no benchmark score can match. Trust is verified, not given. Until then, treat every "available for a fee in the future" as an unpriced liability on your household balance sheet.
The smart home will cost money. That was never in dispute. The question is who holds the meter, and whether you are permitted to read it.