The free tier died on a Monday, and it died with a timestamp.
On October 12, 2026, xAI closed the free consumer path to GrokBot. The base Grok model remains reachable, but the agent layer — the part that plans, calls tools, retries on failure, and actually executes — now sits behind X Subscription Premium at $30 per month. Twenty-four hours later, Apple announced HomeView at a $299 hardware entry price and, in the same breath, disclosed that server-side Apple Intelligence features including Siri AI, photo editing, and Image Playground would carry daily usage limits, with expanded access "available for a fee in the future." Meta had already shown its hand in September: Meta Muse caps the free tier at 100 million tokens per week and sells 500 million for $20, and 3 billion for $100.
Three vendors. Three pricing architectures. One conclusion. There is no free path to a functional home agent, and the industry has collectively decided that the household is a metered account rather than a user.
The free tier did not disappear because compute got expensive. It disappeared because the customer acquisition budget got cut, and nobody in this market is funding your experimentation out of pocket anymore.
I have audited token contracts line by line, farmed yield through a protocol exploit window, and liquidated 80% of my stablecoin book in 48 hours when FTX's ledger stopped reconciling. I know what a pricing page looks like when it is hiding a balance sheet. This one is hiding three, and they are stacked on top of each other.
The Context: What Each Vendor Is Actually Selling
Start with the ledger entries, because the narrative is noise and the numbers are not.
xAI's move is the cleanest signal. A base model stays free. The agent wrapper does not. That distinction matters more than the $30 price point, because it tells you precisely where the vendor believes value accrues. The model is a commodity input. The orchestration layer — memory, tool routing, permissioning, scheduled execution — is the product. xAI is not charging you for intelligence. It is charging you for the right to have something act on your behalf, continuously, inside your home.
Apple's structure is more sophisticated and, in my read, more aggressive. The $299 HomeView is not a subsidy. It is a pump. Apple has never had trouble selling hardware at a margin; what it has historically refused to do is sell a recurring service at a loss. The disclosure that server-side Apple Intelligence features carry daily usage limits, with expanded access priced later, is Apple telling you that the device is the on-ramp and the meter is the business. Note the wording. Not "unlimited." Not "included." Daily usage limits, with a paid tier to follow. That is a rate card published in advance of the rate.
Meta is the most transparent of the three, and transparency deserves credit even when the number is unflattering. Meta Muse launched in September on a pure token ladder: 100 million tokens per week free, then $20 a month for 500 million, then $100 a month for 3 billion. No hardware gate. No ecosystem lock. Just a usage curve with a price attached.
Lay the three side by side and the pattern is not subtle. Runtime access is being gated, hardware is being positioned as the funnel, and token volume is being converted into a monthly bill. What none of them have published is a cost basis. That is where the analysis has to start, because a price without a cost curve is a marketing document.
The Core: Five Things the Pricing Pages Do Not Say
One: The Ladder Is Not a Cost Curve
Run the Meta arithmetic. $20 for 500 million tokens is $0.040 per million tokens. $100 for 3 billion tokens is $0.0333 per million. That is a 16.7% price reduction in exchange for a 500% increase in volume.
Read that again. Six times the volume buys you a sixth of a discount on a per-unit basis.
If these tiers were priced against marginal compute cost, the curve would be steep. Inference cost per token has collapsed repeatedly over the past three years — the direction of travel is unmistakable across every published benchmark and every GPU-hour market I track. A pricing ladder that tracks a deflating input should show aggressive volume discounts at the top end, because the vendor's cost at 3 billion tokens is materially lower per unit than at 500 million. Instead the discount is nearly flat.
That flatness is the tell. The tier structure is calibrated to willingness to pay, not to cost of goods. The $20 tier exists to capture the committed hobbyist. The $100 tier exists to capture the household that has already wired the agent into its lighting, its calendar, and its door locks. Neither price is derived from a cost model. Both are derived from a switching-cost model.
This is the same pattern I documented in 2020 when I decomposed yield farming returns and found that advertised APR was a function of emission schedules rather than underlying cash flow. The headline number told you what the protocol wanted you to do. The cost line told you what was actually happening.
Two: The Token Caps Are Calibrated to Text, Not Vision
Here is where the pricing pages get genuinely misleading, and where my own operational data becomes useful.
A household text agent does not consume much. Model the burn. A moderately active agentic household — say 20 delegated tasks a day, each burning 50,000 tokens across long context, tool-call loops, and retries — consumes roughly one million tokens per day, or 30 million per month. Meta's free tier of 100 million tokens per week equates to about 433 million per month. That is fourteen times the load of a heavy text household.
The free tier, in other words, is not a compute subsidy for realistic text usage. It is a demo. Nobody with a text-only agent will ever touch the ceiling.
Now run the same model with vision. Video understanding is where token consumption detonates. At roughly 258 tokens per frame at one frame per second, a single video stream costs about 15,500 tokens per minute, or roughly 930,000 tokens per hour. Ten cameras running eight hours a day is 80 camera-hours, which is 74 million tokens per day, which is 2.2 billion tokens per month.
Two point two billion. Meta's top tier is three billion for $100. The free tier's 433 million covers about four days of a ten-camera household at half duty cycle.
The caps are not sized to what a household says. They are sized to what a household sees. And vision is exactly what all three vendors are selling hardware to enable. Apple's HomeView is a camera-adjacent hub. Every HomeKit camera you add spins the meter faster. Meta's Muse ladder becomes a genuine constraint the moment you point a lens at your driveway.
That is the mechanism, and it is elegant in the way that only a well-constructed rate card can be. The hardware is the pump. The subscription is the meter. The free tier is sized so that it holds indefinitely for the buyer who never buys the camera, and ruptures within a week for the buyer who does.
Three: The Compounding Stack Is the Real Number
Everyone is quoting $30. The $30 is not the number.
Build the household stack. You have the agent runtime, which is $20 to $30 depending on vendor. You have the cloud model API for anything the runtime does not cover natively, which on the Meta ladder runs $20 to $100. You have the device ecosystem: $299 for HomeView, plus cameras, plus the sensors that make the agent useful in the first place. You have storage for the video the agent is now ingesting, which is $3 to $10 a month at consumer cloud rates and considerably more if you want retention. And you have the skills layer — the integrations, the marketplace plugins, the third-party connectors — which at current market pricing runs $5 to $15.
Stack it honestly. A functional vision-enabled home agent lands between $50 and $150 per month, or $600 to $1,800 per year, before hardware. Over a five-year hardware refresh cycle, that is $3,000 to $9,000 in recurring charges on top of the capital equipment.
For calibration: a typical German household consumes 2,500 to 3,500 kilowatt-hours a year. At current residential rates of roughly €0.35 per kilowatt-hour, that is €875 to €1,225 annually. The agent stack at the top end costs more than the electricity that runs the house.
I am not making a rhetorical point. I am making a budget point. The 'smart' home has crossed the line where its software subscription exceeds its utility bill, and no vendor has published that comparison because it is the single most damaging number in the category.
Consumers evaluate these charges in isolation — $30 feels like a streaming service, $100 feels like a phone plan — and the isolation is the design. Nobody is meant to add the column up. I add columns up for a living.
Four: The Free Tier Was Liquidity Mining, and the Emissions Just Ended
This is the framing that the crypto-native reader will recognize immediately, and it is the correct one.
From 2020 through 2022, protocols paid emissions to bootstrap total value locked. The advertised yields were not returns; they were acquisition costs dressed as returns. When the emission schedules terminated, the mercenary capital left within days, because it had never been there for the product.
Free agent tiers are the identical instrument in a different wrapper. The tokens were never free. They were a customer acquisition cost line item, capitalized against a future conversion rate. When the conversion rate came in below the threshold that justified the spend, the program ended — and it ended with a timestamp, because finance departments do not phase things out gracefully when the number is bad. They set a date.
October 12 is that date.
The lesson from 2021 is that you cannot distinguish a real user from a mercenary until the emissions stop. The lesson from October 2026 is that most free-tier agent users were mercenaries, and the vendors found out before we did. Liquidity vanishes when fear replaces calculation, and so does a free tier when the spreadsheet stops forgiving it.
The bear market context sharpens this considerably. In a risk-off environment, every platform in every category is auditing its loss leaders. The consumer AI agent was the most expensive loss leader in the market, because unlike a freemium SaaS product, the marginal cost of a free user here is not near zero. It is real GPU time. You cannot overprovision your way out of that.
Five: Cost Sensitivity Is a Function of Whether the Agent Has a P&L
Here is the piece of first-person data I can actually contribute, because I have run an agent at scale.
In 2026 I built an automated trading agent executing MEV-resistant arbitrage across decentralized exchanges. It processed roughly 10,000 transactions a day at a 99.9% success rate. At an average of 8,000 tokens per decision cycle across 10,000 cycles, that is 80 million tokens a day, or 2.4 billion per month — sitting almost exactly at Meta's top tier.
At $100 a month, the compute bill for that system was a rounding error against gross captured spread. Sub-one-percent. I never once looked at the token meter, because the agent had a profit and loss statement, and the subscription was a line item inside it.
Now put the same meter in a living room.
A household agent has no P&L. It produces convenience, which is real but unmeasured, and it produces it against a bill that arrives whether or not the convenience materialized. The identical $100 is trivial for the arbitrage desk and extortionate for the family that bought a $299 hub to turn off lights with their voice.
That asymmetry is the entire consumer agent problem, and no pricing page addresses it. The vendors have built enterprise-shaped cost structures and pointed them at households that cannot amortize them against revenue. Until an agent produces measurable income for a consumer — not time saved, income — the subscription will always feel like a tax, because structurally, it is one.
I standardized my agent's code repository so other developers could reproduce the results. What I could not standardize was a reason for a household to keep paying.
Six: A Prepaid Balance Is Unsecured Credit, and the Discount Tells You the Rate
Watch for the annual prepay offer. It is coming, and it will be framed as a saving.
Here is the arithmetic nobody runs. A 17% discount for paying twelve months upfront implies an annualized cost of capital of roughly 20.5% — because you are handing over $830 today to avoid $1,000 spread across the year, and 170 divided by 830 is 20.5%.
Twenty point five percent is a junk-bond rate. That is what the platform is implicitly telling you its capital costs, or what it is willing to pay to pull forward cash in a market where forward cash is scarce. Both readings are credit signals, and neither is a customer-benefit story.
In November 2022 I moved 80% of my stablecoin holdings into non-custodial cold storage within 48 hours of the FTX collapse, because the lesson of that week was not that one exchange was fraudulent. The lesson was that a balance held on someone else's ledger is an unsecured claim, and unsecured claims get haircut in order of arrival. A prepaid annual subscription is the same instrument at smaller scale. You are an unsecured creditor of a service obligation, with no collateral, no priority, and terms that can be amended unilaterally.
Ledgers do not lie, only the auditors do — and a prepay discount is the platform's own auditor quietly disclosing its liquidity preference.
Read the terms of service on the usage limits before you read the price. Apple's daily caps on Siri AI and Image Playground are not written in code. They are written in a support document, which means they can be changed without a software update, without notice, and without recourse. Code executes what lawyers cannot enforce — and where there is no code, there is only a lawyer's document, and the document always favors the drafter.
Seven: Where Crypto Actually Competes, and Where It Does Not
The reflex in this industry is to declare that decentralized payment rails will eat the subscription stack. Most of that argument is wrong, and the part that is right is narrow enough to be actionable.
Start with what is wrong. Agent micropayments do not need a dedicated data availability layer. Run the volume. A busy household agent might settle 400 payment events a day across tool calls, API requests, and skill invocations. At roughly 100 bytes of calldata per event, that is 40 kilobytes per day. Forty kilobytes. A single consumer-grade blob carries more than that in one transaction. The agent payment thesis is real; the dedicated-DA thesis attached to it is a solution in search of a load profile, and the load profile is four orders of magnitude too small.
What is right is narrower and more interesting. The x402-style pay-per-request handshake — an HTTP 402 response that settles a micro-payment inline — solves a problem that card rails genuinely cannot. Card networks have a fixed cost floor per transaction that makes sub-cent billing uneconomic, which is precisely why every consumer agent vendor has retreated to monthly buckets. Buckets exist because the rail cannot meter. A per-request settlement rail does not need buckets.
That is the structural opening. It is not about decentralization as ideology. It is about the fact that a metered product billed through an unmeterable rail has to be sold in prepaid blocks, and prepaid blocks are where lock-in lives. Pair that with the agent identity and registry work around ERC-8004 and its descendants — portable reputation, verifiable execution history, agent-to-agent discovery without a platform in the middle — and you have a coherent alternative: a household agent that pays for exactly what it consumes, carries its memory and reputation with it, and can be ported to a different runtime without re-buying the ecosystem.
Which is, of course, exactly why none of the three vendors on this list will build it. And that brings me to the skill marketplaces, which are the tell for where this goes next. The reason a platform will gate its agent skill catalog is not technical. It is the same reason a game publisher gatekeeps gear drops: the mint is the monetization. The moment a platform can arbitrarily issue and revoke capabilities, it owns the economy. Portability is the enemy of that model, and portability is the first thing to be quietly deprecated.
The Contrarian Read: This Is Not a Compute Price
The consensus explanation for October is that compute is too valuable to give away. I think that explanation is comfortable, convenient, and wrong on the arithmetic.
Inference cost per token has been falling, not rising, across every credible benchmark I have seen for three consecutive years. Hardware efficiency improves. Serving stacks get better at batching and speculative decoding. The marginal cost of a text token in 2026 is a fraction of what it was in 2024. If the constraint were compute, consumer agent prices should be falling, and the free tier should be expanding.
Instead the free tier contracted on a specific date and the ladder came in nearly flat.
So the price is not a compute price. It is a distribution price, and it is set by switching cost. The number a vendor can charge is the number it would cost you to leave — the cameras already mounted, the routines already scripted, the family already trained on the wake word. That number has nothing to do with GPU-hours and everything to do with how deeply the agent is threaded into your home.
The uncomfortable corollary: consumers should prefer metered pricing to "unlimited." Unlimited is where the throttling hides. A published rate card with a visible ceiling is a contract you can audit. An unlimited plan with an undisclosed fair-use policy is a discretionary promise, and discretionary promises get repriced quietly, mid-term, with a support article. Give me the meter. I can do math on a meter.
We trade the protocol, not the promise. The meter is the protocol.
Takeaway: What to Watch, and What to Do With Your Balance Sheet
Three signals matter from here.
First, whether a spot market for agent compute emerges on-chain. Tokenized GPU-hours with a published clearing price would immediately expose how far the consumer ladder sits above cost, and that exposure is the single largest threat to the current pricing architecture. Watch for the first credible benchmark, not the first token launch.
Second, the prepay offers. When they arrive — and they will, in the fourth quarter, when platforms need cash — read the discount as a credit spread rather than a saving. Anything above 15% annualized is the platform telling you something about its own balance sheet that its investor deck does not.
Third, the portability question. Ask, before you mount the camera, whether your agent's memory and skills can leave with you. If the answer requires reading a support article, the answer is no.
The free tier died on October 12. The question worth asking is not who killed it. It is why the industry needed a specific date to admit that the household was always the product — and how much of your living room you have already pledged as collateral.
