This report began as a discussion between two builders. Both agreed the moment is extraordinary. Both were anxious. But they were anxious about different clocks. One believes the window is closing: today's AI is subsidized, underpriced, and temporary — exploit it before the bill arrives. The other believes the floor is falling: every generation of capability gets radically cheaper, and the real edge was never the price. The data, it turns out, has something to say to both of them.
The feeling has a name in the literature now. Researchers call it FOMO-AI — the worry that your AI skills, access, or output lag behind others — and it has a validated psychometric scale. In a 557-person U.S. study, elevated FOMO-AI predicted greater anxiety and depressive symptoms, with younger adults and women disproportionately affected. One finding stands out: AI literacy buffers the effect. The people who understand the tools best worry the most productively — and the least destructively.1516
Not generalized tech anxiety — a specific, measurable fear of being left behind by AI. It correlates with anxiety and depressive symptoms, which in turn reduce well-being. This is the first peer-reviewed anchor for what builders have been describing anecdotally since 2023: staying up until 3 a.m. because the tools feel too good to put down, and too temporary to trust.15
Strip away the specifics and almost every AI-literate builder is running one of two mental countdowns. Both are rational. Both are backed by real evidence. They cannot both be the clock that matters most.
AI capability is massively ahead of its price. For a few hundred dollars a month, one person can produce the output of an extremely capable team. That gap exists because labs are burning investor capital to win users — the Uber playbook: subsidize until habits form, then charge what it actually costs.
If that's true, this is a brief, historic arbitrage. The people who build aggressively inside the window will look like the people who registered domains in 1996. Everyone else will look back with regret.
The moment is real — but the pricing premise is backwards. Every model generation gets cheaper per unit of capability, at rates with no precedent in any prior technology. Today's frontier model becomes tomorrow's cheap baseline. The premium tier may stay expensive; the floor keeps dropping.
If that's true, the window doesn't slam shut — access keeps broadening. The durable edge isn't getting in before the prices rise. It's taste, judgment, and knowing where premium intelligence is actually needed.
Both clocks were wound by the same events. Read the timeline once through the Arbitrageur's eyes, then again through the Compounder's. It supports both stories — which is exactly why the anxiety won't resolve on its own.
ChatGPT launches. Querying a model at GPT-3.5-class performance costs about $20.00 per million tokens. Nobody calls this expensive — there is nothing to compare it to. The clock starts.1
GPT-4-class models arrive at premium prices, and the first builder migration begins. The pattern that will define the era emerges immediately: the newest capability commands a steep premium, while last year's miracle quietly gets discounted.
Price declines accelerate. Across benchmarks, the cost to hit a fixed capability level falls between 9× and 900× per year depending on the task — and after January 2024 the median rate accelerates from roughly 50× to roughly 200× per year. No prior technology curve looks like this.34
Open-weight models close the quality gap with closed models from 8% to 1.7% in a single year on human-preference leaderboards. By late 2025 they serve roughly a third of all tokens on OpenRouter, with Chinese open-weight models alone surging from a 1.2% weekly share to nearly 30% in some weeks. Capability is leaking out of the labs' pricing power.15
The other clock chimes. Replit replaces flat $0.25-per-checkpoint Agent pricing with effort-based pricing that scales with AI consumed — bills spike, users revolt.6 Weeks later, Anthropic introduces rate limits to curb Claude Code power users.9 The flat-rate, all-you-can-eat phase of AI products is visibly ending.
Reasoning models — which burn far more tokens per task — grow from negligible to more than half of all tokens routed on OpenRouter. Unit prices keep falling; total consumption rises faster. The bill goes up anyway.5
Reporting on Anthropic's internal figures shows a gross margin of −94% in 2024 swinging to a projected ~+40% for 2025 — undercut by inference costs running 23% over plan. The labs really were losing money serving you. And the loss really is shrinking. Both builders claim vindication.78
The Arbitrageur's argument rests on a simple suspicion: nobody sells dollars for ninety cents forever. If the current price of intelligence is a customer-acquisition expense, then today's builders aren't customers — they're the habit being formed.
Today's AI pricing reflects what labs are willing to lose, not what intelligence costs. Venture and investor capital absorbs the gap while usage habits form. Once dependence is established and capital tightens, prices converge upward toward true cost plus margin — and the builders who structured their economics around subsidized intelligence get repriced overnight. It is the Uber playbook applied to cognition: cheap rides until you've sold your car.
Serve intelligence below cost. Anthropic's gross margin in 2024: −94%. Every token you bought, someone else partly paid for.
Builders restructure entire workflows, products, and businesses around the subsidized price. The dependency is the product.
Flat rates become usage rates. Checkpoints become "effort." Unlimited becomes rate-limited. 2025 provided the first documented wave.
With habits locked in and alternatives behind, prices settle at what the market will bear — which, for the dependent, is a lot.
Unaudited internal projections reported by The Information via Axios (March 2026). The 2025 projection was itself cut roughly 10 points because inference costs on Google and Amazon clouds ran ~23% higher than planned, with 2025 inference spend growing more than threefold to roughly $2.7B. Note what the Arbitrageur must concede: this chart shows serving economics improving dramatically — the strongest version of the window thesis is about product pricing, not raw model economics.
Before: Replit Agent charged a flat $0.25 per "checkpoint" — a predictable, arguably subsidized unit price for agentic work.
After: "Effort-based pricing." Complex tasks bundle into a single checkpoint that can cost substantially more than $0.25, scaling with the AI effort consumed. Users documented bill spikes; the backlash was loud enough that Replit acknowledged rollout problems and issued credits.6
Weeks later, Anthropic imposed rate limits on Claude Code's heaviest users.910 Cursor's 2025 pricing changes triggered the same cycle of confusion, anger, and refunds.11 One quarter, three repricings — each moving the same direction: toward usage-aligned billing.
Sustained gross-margin improvement at the labs without corresponding end-user price increases; flat-rate products that survive their power users; per-task costs (not just per-token prices) continuing to fall through 2026–2027. The margin chart above is already uncomfortable evidence: the subsidy is shrinking because serving got cheaper, not because prices went up.
The honest version of the window thesis survived verification in a narrower form than its believers tell it: product-level subsidies are demonstrably ending — flat rates, unlimited tiers, loss-leader checkpoints. What did not survive is the claim that intelligence itself is about to get more expensive.
The Compounder's argument doesn't need a story. It needs a chart. The price of any fixed level of AI capability has collapsed at rates with no precedent in the history of technology pricing — faster than Moore's Law, faster than solar, faster than genome sequencing.
Whatever the frontier costs, last year's frontier becomes this year's commodity. Smaller models hit the same benchmarks, hardware gets 30% cheaper per unit of performance every year, and open-weight models put a hard ceiling on what anyone can charge for mid-tier capability. The premium tier may stay premium — but the capability most execution work actually requires keeps falling toward free. The window doesn't close. The floor drops out from under it.
A more than 280-fold price reduction for equivalent benchmark performance, driven by increasingly capable small models. Bar lengths are illustrative — at true scale, the second bar would be less than half a pixel wide. Source: Stanford HAI 2025 AI Index, data from Epoch AI. Figures are API list prices at MMLU-equivalence, not raw compute cost.
Epoch AI, March 2025: log-linear regression of the lowest-priced model meeting fixed performance thresholds across six benchmarks, ~2022–early 2025. Bar lengths are proportional to the logarithm of the decline rate. Epoch cautions the fastest recent rates may not persist, and the analysis predates reasoning-model economics. Underneath it all: ML hardware costs fall ~30% per year performance-adjusted, and energy efficiency improves ~40% per year.1
By late 2025, open-weight models served roughly one-third of all tokens on OpenRouter — with Chinese open-weight models alone growing from a 1.2% weekly share in late 2024 to nearly 30% in some weeks.5 Whatever closed labs decide to charge, this is the ceiling on mid-tier pricing power. Caveat: the gap figure is a dated snapshot on a human-preference leaderboard, and OpenRouter telemetry skews toward cost-sensitive routing traffic.
Price-per-fixed-capability flattening or reversing across 2026–2027; open-weight models failing to track the frontier as reasoning workloads dominate; or per-task costs rising persistently because capability gains demand disproportionately more tokens. Epoch's own caveat applies: the historic rates were measured before reasoning models changed what a "task" consumes.
The floor thesis won the verification battle decisively on unit economics. But notice what it quietly assumes: that the capability you need tomorrow is the capability that's getting cheap today. That assumption is where the third clock starts ticking.
In 1865, William Stanley Jevons noticed that more efficient steam engines didn't reduce England's coal consumption — they exploded it. Cheaper intelligence is doing the same thing. As unit costs collapse, builders don't pocket the savings. They upgrade to more compute-hungry capability.
Reasoning models consume far more tokens per task — they think out loud, at length, before answering. As they became the default for serious work, total token consumption surged even as per-token prices fell. OpenRouter data represents one API marketplace skewed toward developer workloads, not the whole market.
A roughly 240× spread between frontier and budget pricing persisted through 2025 — even as both tiers got cheaper in absolute terms. Bar lengths are illustrative, not to scale. This spread is precisely what underwrites the working pattern both builders already converged on: frontier intelligence for planning, architecture, and first-pass thinking; cheap models for execution and edits.
This is the synthesis hiding inside the argument: both builders are describing the same curve from different ends. The Arbitrageur watches total spend rise and calls it the subsidy ending. The Compounder watches unit prices fall and calls it the floor dropping. Jevons says: yes. Both. That's what the curve does.
Total spend can rise while unit cost collapses — when falling prices unlock demand faster than they save money.
Both clocks assume the race is crowded. The adoption data says otherwise. The feeling that "everyone is already ahead" is an artifact of who builders follow, not of who actually uses these tools. Most of the labor market hasn't seriously started.
Federal Reserve Bank of St. Louis / Real-Time Population Survey (Bick, Blandin, Deming). Personal use is sprinting; work use is crawling — roughly 63% of U.S. workers still don't use generative AI at work at all. Gallup's Q4 2025 wave shows workplace adoption roughly plateauing at 46%, with 49% answering "never."13 Habitual use is rising (28% weekly, 13% daily by Q1 2026) but remains a minority behavior.
Sit with the asymmetry: builders are having a sophisticated argument about which decimal of the token price matters, while half the working-age population has never opened the tools at all. Whatever the right clock is, the stadium is mostly empty. The race has barely started — and the anxiety of the people in it is calibrated to the wrong crowd.
Put both clocks on the table and the disagreement gets surprisingly small. They aren't predicting different futures. They're watching different layers of the same stack — and each layer has its own expiration behavior.
| The Closing Window | The Falling Floor | |
|---|---|---|
| What it watches | Product pricing — subscriptions, flat rates, agent checkpoints, usage tiers. | Capability pricing — the cost of hitting a fixed benchmark via API. |
| What the data shows | Real and ending: Replit's effort-based repricing, Claude Code rate limits, Cursor's billing revolt. Flat-rate AI was a loss leader, and 2025 was the year it died.6911 | Real and accelerating: 280×+ collapse in two years, ~200×/yr median decline post-2024, open weights one model-generation behind.13 |
| What actually expires | The deal — unlimited tiers and below-cost products. Not access to intelligence itself. | The premium — on any fixed capability level. The frontier stays expensive; everything behind it commoditizes. |
| The blind spot | Confuses the end of product subsidies with the end of cheap intelligence. The margin data shows serving costs falling, not prices rising.7 | Assumes per-token price is the bill. Jevons says the bill is tokens × price — and tokens are exploding faster than prices fall.5 |
| Rational response | Build aggressively now — but structure unit economics to survive usage-aligned pricing, because that's what's coming. | Invest in taste, judgment, and model-routing skill — knowing where frontier intelligence is needed and where the cheap baseline suffices. |
The verified record, compressed: the window thesis is right about products and wrong about intelligence. The floor thesis is right about intelligence and silent about bills. The anxiety is real, validated, and aimed at the wrong target — the scarce resource was never this year's pricing. It's the judgment to know what to build while everyone else is still deciding whether to show up.
Which clock are you building against?
Two builders argued about which clock was real. The data's answer is less dramatic and more useful: the subsidized deal expires, the price of intelligence keeps falling, and your total bill rises anyway — while half the market hasn't entered the building. The anxiety is measurable, the window is mislabeled, and the moat was never the discount. It was knowing what to do with the machine.
Methodology note: every quantitative claim in this dossier passed a three-vote adversarial verification process against primary sources. Figures that failed verification — including a widely circulated OpenAI burn projection and several token-share statistics — were excluded. Anthropic margin figures are unaudited internal projections relayed through secondary reporting and are flagged as such. Epoch AI's decline rates measure price-to-fixed-capability (API list prices), not frontier-model pricing or raw compute cost.
