Token Costs Are Collapsing: The Flat AI Subscription Will Fall

The thesis: flat pricing is a political price with an expiry date

The flat-rate AI subscription is a political price. The variable that matters is the cost of inference per million tokens, and that variable writes its expiry date.

On 5 October 2026 SemiAnalysis published the measure that holds up everything else. Subscriptions account for roughly 10% of Anthropic's revenue and absorb over 40% of inference compute (SemiAnalysis, Tokenomics Model, 5 October 2026^[1]). The same calculation estimates a cut in average revenue per megawatt of close to 36 million dollars.

That study also measures a gap of more than 5 times in value per dollar between Anthropic's plans and OpenAI's. The news travelled across the technical press on 6 October (The Register^[2]).

This is a regime change, far more than a trend. The flat fee will fall by arithmetic, and the arithmetic is already written into public price lists.

The consensus has the wrong frame

The consensus reads that gap as a generosity contest. Whoever gives away more tokens wins the customer, and the ranking shifts with every model release.

The consensus has the wrong frame. A 5-times gap at the same list price measures something else: the distance between the retail price and the marginal cost of serving the tokens.

When two vendors ask 200 dollars a month for quantities of work that differ fivefold, that price carries zero information about cost. It carries information about the customer acquisition budget. It is a marketing line dressed up as a price list.

The figure everyone watches is the monthly cap. The figure that predicts is the ratio between credit burned and the API price of the same workload, and the 5 October source shows that the two ratios diverge sharply.

Ninety per cent of analysts read the present well. They get the tempo of change wrong.

The curve: three points in dollars per million tokens

A trajectory needs three points. This desk keeps a dedicated accounting class, with value, date and measurer for every point in dollars per million tokens.

First point, from this desk's archive: between March 2023 and the summer of 2026 the price per million tokens for GPT-4-level capability falls roughly a thousandfold, read off public API price lists.

Second point, same series: at fixed capability the price drops about 40 times a year. And demand concentrates, given that 29% of production tokens run on just 4% of the spend.

Third point, carried by the 5 October 2026 source: the API-equivalent value of the same 200-dollar plan shifts by model and by workload type, and the ratios between credits diverge from the ratios between API prices.

Three points compose a single direction. The unit cost of a token collapses, the quantity consumed per user climbs faster than the price falls, and the flat fee absorbs the difference.

The mechanism: credits turn a fixed price into a variable quantity

The monthly payment buys credits. Each pairing of model and token type burns a different share of those credits.

Here sits the causal mechanism, far more than a correlation. The vendor fixes the price and leaves the quantity variable, then recalibrates the credit cost when margins tighten. Public limits change with promotions; credit costs change in silence.

This is already a meter. What it lacks is the label.

The source says it explicitly: modelling an AI lab's accounts requires understanding subscription limits, day by day. SemiAnalysis has built a dashboard that re-reads every combination of plan, model and token type, with Meta, Cursor, Cognition, Z.ai, MiniMax and Moonshot alongside the two giants.

A hidden meter turns fragile the moment someone publishes it. Publication arrived on 5 October 2026.

Cliff event: 2027, when the yardstick goes public

Adoption of a new price jumps, far more than it grows by degrees. The jump arrives when the measure becomes public and comparable across vendors.

Cliff event: the dashboard makes every credit recalibration visible within hours, silent tweaks become news, and the vendor prefers a declared usage-based price list. Horizon: 2027.

The mechanics stay simple. A downward tweak today costs reputation, given that Anthropic has backed off planned cuts more than once to avoid the court of public opinion. An explicit price list costs once and that is it.

The bottleneck helps the transition. Scarcity migrates from training silicon to packaging and memory, and whoever prices it as permanent loses: the same publisher maintains a model dedicated to Chinese data centres, where cost per megawatt remains the central variable (SemiAnalysis, China Datacenter Model^[3]).

Capacity arriving plus unit cost collapsing: the subsidy loses its commercial function.

Three categories that change shape by 2028

• Flat-rate coding agent resellers (Cursor, Cognition): the margin depends on credits bought from third parties and resold at a frozen price.
• Unlimited plans at vertical SaaS companies: whoever includes inference in the fee inherits the volatility of cost per token.
• Labs with aggressive price lists (Z.ai, MiniMax, Moonshot): they use price as an entry lever, and they gain when the market moves to metering.

The first group carries the worst risk. Unit cost falls, and agentic workloads raise tokens per task on a steeper gradient: the final balance depends on an accounting decided by others.

The second group will discover that inference included in the fee is a derivative, far more than a product feature.

The third group gains from the shift to metering, because a common yardstick rewards the lowest cost per token. A fourth effect concerns the buyer: the solo operator who arbitrages flat pricing today shifts the advantage to the local model, where cost per token stays fixed and known in advance.

The position, and the evidence that overturns it

My position: the flat price of AI subscriptions will converge towards the marginal cost of inference, and the shift will happen by arithmetic, with declared usage-based price lists in place of opaque caps.

The reasoning runs in three steps. The cost per million tokens has been falling on a measurable gradient for three years. Subscriptions consume compute far more than proportionally to the revenue they bring, and that gap has become public and updated every day.

What would change my mind: a consumption-per-user curve that runs faster than the descent of unit cost for two years running. In that world the flat fee remains the most efficient form of rationing, and the explicit meter becomes a pointless expense.

What to do now, by role:

• CTO: reassess the stack built on frozen API prices, and measure cost per task in place of cost per user.
• Venture capital: the bet that looks impossible is the vertical that buys tokens on metered pricing and sells outcomes, with a moat on data.
• Chief strategy officer: a three-year plan built on a fixed fee per user describes a world on its way off stage.
• Procurement: insert a repricing clause tied to dollars per million tokens, with a review every six months.

The price of intelligence falls every quarter. Contracts signed now last three years.

Prediction, horizon, kill signal

Prediction: by 31 December 2027 at least one of OpenAI and Anthropic replaces its flagship consumer flat plan with a declared usage-based price list, or publishes a cap measured in tokens within the price list.

Confidence: 70. Horizon: 31 December 2027. Kill signal: at 31 December 2027 both vendors hold flagship flat plans at an unchanged price and with limits equal or more generous, and the SemiAnalysis subscription dashboard records API-equivalent value per dollar rising over twelve months.