The most interesting number in AI pricing right now is not a price. It is the gap between two directions that are moving at once. Consumer plan prices have crept up and fragmented into tiers that run from free to $200 a month, while the price of the thing those plans are made of, the tokens, has collapsed. Between June and October 2026, the two largest US labs cut model prices by as much as 80%. My read is that the labs have quietly stopped selling you intelligence and started selling you certainty: a flat monthly fee that caps your exposure to a commodity whose cost is falling under their feet. The winners of this phase are not the companies with the best model. They are the ones who have figured out who to charge, how, and for what.
The headline price and the real price are moving apart
Start with what a buyer actually sees. Anthropic now runs four consumer tiers: free, Pro at $20, Max at $100 and Ultra at $200, with Ultra being the plan it used to sell as Max 20x before a late-October rename. Since 7 October 2026, Max and Ultra also hand back a monthly API credit that matches a chunk of their own price, $100 on Max and $200 on Ultra. OpenAI mirrors the structure with $20, $100, $200 and $500 tiers, and Google, xAI and Perspectivity fill in around them.
Underneath those sticker prices, the cost of the raw material fell off a cliff. Anthropic made Claude Sonnet 5's introductory rate of $2 per million input tokens and $10 per million output tokens permanent in August, then answered OpenAI's September cuts to its GPT-6 line with Opus 5.5 at $4/$20 list, carrying efficiency gains the company put at roughly 40% in effective savings. OpenAI halved API prices on its mid-tier models in September.
So the headline prices held or rose while the unit cost sank. That spread is the whole game. A flat subscription is a bet by the provider that it can serve your usage for less than you pay, and the falling token price keeps widening the margin on that bet, as long as you do not use the plan to its limit. The clever twist is that the providers quietly adjust the limits rather than the price. One analyst flagged that OpenAI halved the Codex allowance on its $200 tier the day a cheaper model launched. The price on the box stayed the same. What you got for it shrank.
Flat fees versus the meter, and why Anthropic chose the meter
There are two ways to charge for AI, and the choice reveals what a company believes about its own economics. Flat-rate subscriptions dominated the early years, with fixed monthly fees for generous or effectively unlimited access. The meter, per-token billing, ties revenue to actual consumption.
Anthropic has been moving enterprise customers off flat rates and onto per-token billing, so the revenue it books reflects what customers really burn. Its new enterprise plans charge per seat with token usage billed on top at API rates. This looks like a minor billing detail and is actually a strategic statement. Flat rates were leaking money: developers were routing cheap conversational subscriptions through agentic tools that ran models around the clock, and a heavy Claude Code user on a $200 Max plan could be pulling usage that would have cost up to $5,000 at list. Per-token billing closes that arbitrage and, as one observer put it, gives Anthropic cleaner data on what customers actually value. OpenAI, whose model leans harder on maximising token spend, ends up with bigger numbers that are harder to prove are real.
There is a deeper point here that most coverage skips. Subscriptions are a small slice of these companies' revenue but a large slice of their compute. On SemiAnalysis's rough figures for Anthropic, subscriptions are around 10% of revenue yet can consume over 40% of inference capacity, dragging blended revenue per unit of compute down by tens of millions. Flat-fee power users are the least profitable customers these businesses have. The migration to the meter is, in part, a quiet effort to make those users pay their way.
The margin reality nobody puts on the pricing page
For a buyer trying to judge who holds the power, the number that matters is gross margin, and it is nothing like classic software. Traditional SaaS runs at 75% to 85%. AI-native products are projected to average around 52% in 2026, up from 41% in 2024, with inference making up roughly a fifth of product cost. ICONIQ's work shows the trajectory improving, from 45% in 2025 to a projected 53% in 2026 and 59% in 2027, but still well short of the software companies investors are used to underwriting.
Pricing model drives a surprising amount of this. Usage-only pricing posted the lowest median gross margin in one 342-company benchmark, at 62%, against 76% to 84% for subscription variants. In other words, the meter protects the vendor against runaway usage but can compress its margin if the model is priced thin, while subscriptions do the opposite. The best operators are learning to run both at once: a flat fee to anchor the relationship and set expectations, with metered overage to catch the heavy users before they turn into a loss.
The raw-material picture is genuinely improving for the labs. SemiAnalysis reported that Anthropic's inference infrastructure margins rose from 38% to over 70% as its annual revenue ran from around $9B toward $44B. Falling token prices are partly a pass-through of that efficiency and partly a weapon aimed at open-source alternatives, cutting rates far enough that self-hosting stops being worth the trouble.
The demand side just blinked
Here is the development that should temper any straight-line optimism. Ramp's data shows token spend grew just 1.5% across August and September, after growing 42% month-on-month in July. Usage kept climbing; spending did not. The price cuts and smarter cost controls among the biggest buyers have, for now, capped revenue even as adoption spreads, with around 56% of US businesses paying for AI.
Anthropic itself said in a filing that its price reductions on Fable 5.1 and Opus 5.5 trimmed revenue growth by roughly $600 million. That is a company choosing share and stickiness over near-term revenue, right before an IPO, which is an unusual thing to put in a prospectus and tells you how seriously both labs take the land-grab.
What it means
The market read is that pricing power in AI is migrating away from raw model access and toward the layer that packages it. The model is becoming a commodity whose price only goes one way. Value is accruing to the subscription relationship, the workflow a tool is embedded in, and the proprietary data and distribution around it. When you assess an AI company now, the question I would push hardest on is gross margin and its direction, not model benchmarks. A firm priced purely on usage with a thin model margin is far more exposed than one with a subscription anchor and metered overage.
Watch three things. The spread between sticker price and unit cost, because that is where margin lives and it is currently widening in the labs' favour. The quiet erosion of subscription allowances, which is a price rise dressed as a product update, and the honest signal to track instead of the headline fee. And the demand stall in the Ramp data, which is the first real evidence that cheaper AI may not translate into proportionally more revenue, an uncomfortable fact for anyone underwriting these IPOs on hypergrowth.
If you are a buyer, the practical move follows from the same logic the newsletters landed on: most of the value sits in the $20 tier, and a flat plan beats the meter as long as your work fits inside its limits. The moment you run agents daily or build on the API, re-price yourself, because that is exactly the usage the providers are redesigning their plans to stop subsidising.
The companies that profit in this phase are not selling the smartest model. They are selling the bill you can predict, on economics the buyer cannot see. Knowing that the real price is falling while the one you pay holds steady is the single most useful thing to carry into any AI purchasing or investment decision this year.