Developer Theo Browne posted a video commentary this week flagging what he characterised as a deteriorating usage experience on Grok’s higher-tier subscription. The evidence he presents is a single data point relayed from someone else — not his own usage, not a controlled measurement. His account: a user “did $40 of usage and it burned 8% of his weekly limit.” That is one person’s figure at one moment, passed through one intermediary. Everything that follows in this article treats it as exactly that: an anecdote, not a statistic.
Browne’s characterisations — that the $300-a-month plan was previously “basically unlimited Grok,” that limits “now actually feel bad,” and his verdict “not looking great here” — are his own assessments of subjective experience rather than published terms or documented changes. He also references a figure for what he says competing $200 plans offer, but the unit attached to that figure is never stated in the clip, so this article does not repeat the number and does not name the plans in question. Reproducing an unanchored figure would add precision without adding meaning.
One further thing this article will not assert: nothing in Browne’s video, and nothing elsewhere in the public record cited here, establishes whether any limit, policy or plan term was changed by xAI. An allowance can feel smaller for two entirely different reasons — the cap moved, or consumption per task rose — and the available evidence does not distinguish between them. Both possibilities remain open.
The key insight: A metered subscription has a sticker price that is visible and an effective price that is not. The effective price is denominated in tasks completed — and it can move without anyone touching the sticker, the plan name, or the published terms.
The Structural Read
This is the third pricing channel worth naming together, because the conjunction is where the structural insight lives. The first channel is the posted token rate — a number that can hold flat while model capability rises, making the effective price per completed task appear to fall. The second is consumption per task — how much of a vendor’s own meter a single request burns, a figure no vendor publishes and which has to be inferred by third parties. The third is the allowance under a cap, which is where the same underlying dynamic lands for subscription buyers instead of pay-per-token buyers.
Put those three channels together and the asymmetry becomes structurally legible. A single shift in how much a model consumes per request can simultaneously surface as an apparent price reduction in one channel and as a quota reduction in another. Neither movement appears on any price list. The published price moved once, in one direction, for one group of buyers — while the effective price moved twice, in opposite directions, for two different groups. That is not a claim about Grok specifically. It is a property of how metered subscriptions interact with variable consumption, and it applies to any capped AI product.
Harness Theory — Structural Property
The Dual-Price Architecture of Capped AI Subscriptions
Every capped AI subscription carries two prices. The sticker price is what you pay per period. The allowance price is what that payment buys, measured in tasks completed — and it is set by the intersection of the cap and the per-task consumption rate, neither of which is published together on any pricing page. Companies that harness AI through subscriptions inherit this architecture whether they intend to or not. Understanding it is a prerequisite for evaluating any capped plan on any platform.
This also explains why the obvious calculation — dividing $40 by 8% to derive an implied weekly allowance — is one this article explicitly declines to perform. That derived figure would inherit every weakness of its inputs: one user’s workload relayed at second hand, a percentage that is presumably rounded, no established baseline for whether the workload was typical, and a comparison figure with no stated unit. A derived number looks considerably more precise than any of its sources, which is exactly the mechanism by which a single screenshot becomes a widely cited statistic. The same standard this publication applies to vendor figures applies here.
Three Implications
FOR SUBSCRIBERS: The only reliable personal measure is your own repeated task
A headline plan price cannot tell you what the plan buys in practice. A benchmark cannot either — benchmarks measure quality, not consumption. Another user’s screenshot cannot tell you, because their workload is not yours. The only signal that holds is tracking how much of your own allowance a known, repeated task consumes over time. That holds the task constant and lets the consumption rate move — which is the only way to observe the variable that actually matters.
FOR VENDORS: Pricing transparency has a structural gap no pricing page currently fills
The missing disclosure is not the cap — that is usually published. It is the per-task consumption rate, which determines what the cap is worth in practice. Without that figure, a posted allowance is a number without a denominator. Vendors that make this legible — even approximately — change the information environment for buyers without necessarily changing the plan itself. Whether any vendor chooses to do so is an open question; this article takes no position on it.
FOR ANALYSTS: The anecdote-to-statistic pipeline deserves its own scrutiny
Browne’s video will circulate. The implied weekly allowance figure — the one this article refused to compute — will be computed by others and will travel further than the caveats attached to it. That is not a criticism of Browne, whose framing was appropriately hedged. It is an observation about how derivative figures acquire authority as they propagate. The structural property described here is real and worth understanding regardless of whether the anecdote that surfaces it is representative.
Business Engineer Framework
Harness Theory: Understanding Where You Sit in the AI Stack
The dual-price architecture of metered subscriptions is a Harness Theory problem: companies and individuals that consume AI through subscription wrappers are exposed to pricing dynamics that live below the surface of any advertised plan. The Map of AI maps the full stack — from infrastructure through to the subscription layer — and shows where each dynamic originates and where it lands.
One user’s anecdote — relayed at second hand, with no established cause and a comparison figure missing its unit — is thin evidence for any claim about any specific plan. But the structural property it points toward is solid and underappreciated: every capped AI subscription carries a second price that no pricing page publishes, denominated in tasks rather than dollars, and movable without anyone touching the advertised terms. That gap exists on every platform. The anecdote is not the story. The architecture is.
Sources: Theo Browne, YouTube video commentary (September 2026). Structural analysis is original to FourWeekMBA / Business Engineer. No company statements, published plan terms, subscriber figures or revenue data were available to this article; no assertion is made that such information does not exist. This article does not constitute investment or purchasing advice.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
The usage figures above are one individual’s, relayed at second hand in a video commentary — an anecdote, not a measurement. The comparison figure offered for other providers has no unit stated anywhere in the source, so it is not interpreted here, and no other provider’s plan terms appear above. No cause is established: an allowance can shrink in effect because a limit changed or because consumption per task rose, and nothing above states or implies that any company changed, cut or tightened any plan, limit or policy. No implied allowance has been derived from those figures, deliberately — a derived number would inherit every weakness of a single rounded anecdote while looking far more precise than it is. The descriptions of the earlier plan and of the situation generally are the speaker’s characterisations. Nothing above calls any plan bad value, recommends any product or vendor, or advises anyone to switch anything. This is not investment or purchasing advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
