Model Prices Have Fallen, but Key Evidence on Investment Recovery Is Still Missing

The Price Cuts Are Clear, and So Is Their Scope

The most direct change in these materials concerns model pricing. Anthropic announced that, compared with Haiku 4.5, Claude Haiku 5.5 prices were reduced by 90% for requests of no more than 100,000 tokens and by 50% for longer requests. OpenAI announced price cuts of 80% for GPT-5.6 Luna and 20% for Terra. These announcements support a limited conclusion: the cost of using the specified models has fallen. But the reductions vary across products and request lengths. The largest reduction cannot be treated as the change in a company’s overall average selling price, much less as evidence that revenue across the entire industry is shrinking.

Evidence: [3] [4]

Usage Growth Must Pass Two Tests

The WeChat account argues that increased usage does not guarantee recovery of investment in compute. This distinction is worth preserving. It invokes the Jevons paradox to explain how lower prices might stimulate usage; a necessary premise of this explanation is that demand is sufficiently sensitive to price. My analytical judgment here is that, even if this mechanism holds, we must first test whether additional paid usage can offset the decline in unit prices, then test whether revenue can cover service costs and capital investment. The available pricing announcements confirm price changes but do not provide the comparable data needed for either test. Possible usage growth therefore explains a mechanism; it is not yet evidence of corporate profitability.

Evidence: [3] [4] [5]

Why Prices Fell Still Has Several Possible Explanations

The WeChat account interprets Anthropic’s price adjustment as a reactive move to follow competitors. The official pricing materials do not establish this motive. OpenAI’s release materials link efficiency gains to pricing changes, but that likewise cannot substitute for corporate profitability data. Several explanations can therefore be proposed for verification: lower prices may reflect improved service efficiency, an effort to attract demand, or both. These are inferences. If unit service costs fall at the same time, price cuts may not squeeze profits. Pressure to recover investment may increase if costs do not fall enough and paid usage does not grow sufficiently. It is currently impossible to determine which scenario predominates.

Evidence: [3] [4] [5]

Hardware Results Cannot Fill the Gaps in Model Companies’ Accounts

TSMC’s consolidated revenue in September 2026 fell 0.6% month over month and rose 54.6% year over year. The same announcement reports both a month-over-month decline and year-over-year growth. The former alone cannot establish a decline in AI demand, and the announcement does not explain the reason for the month-over-month change. Samsung’s third-quarter 2026 consolidated sales of approximately KRW 195 trillion and operating profit of approximately KRW 107.4 trillion are earnings guidance that has not yet undergone an external audit. The materials did not verify analyst expectations, so these figures cannot be described as an earnings miss. These data describe hardware companies’ consolidated operating results; they cannot directly answer whether model API revenue covers compute costs.

Evidence: [1] [2]

Conclusions Should Go Only as Far as the Evidence Allows

My judgment is that model price cuts make the test of investment recovery more concrete: we need to observe paid usage, revenue, and service costs on a consistent billing basis, along with the payback period for capital investment. Price cuts for the specified products are supported by the evidence, while changes in revenue and costs across the full product portfolio remain unclear. The available evidence is insufficient either to conclude that AI has failed to establish a financially sustainable commercial cycle or to guarantee that lower prices will ultimately improve profitability. Only if comparable revenue and cost disclosures become available can we assess further whether additional demand has produced sustainable improvements in operating performance. Neither hardware revenue nor the price of a single model can establish this on its own.

Evidence: [3] [4] [1] [2]

What to watch next

• Whether products with price cuts disclose paid usage and revenue on a consistent billing basis. Whether additional demand can offset lower unit prices.
• Whether model companies’ unit service costs fall at the same time. Whether comparable data support changes in gross profit after the price cuts.
• Whether verifiable disclosures of capital investment, compute commitments, and payback periods become available, allowing operating revenue and long-term expenditure to be compared on a corresponding basis.