Thesis · Foundation Models & AI Intelligence
Will the per-million-output-token list price of the flagship frontier models from OpenAI and Anthropic fall by at least 40% from their 2025 levels by the end of 2027?
Competition from near-frontier open-weight models (especially Chinese labs) and zero-switching-cost routing forces US frontier labs into repeated list-price cuts, so flagship output-token prices fall at least 40% versus 2025 levels.
Who is on each side
The two sides.
Each side in DeepStack’s words, then the voices who argue it, attributed and paraphrased, dated where the date is known. A voice links to its page on DeepStack where one exists.
The case for yes
Model prices fall toward cost
Near-zero switching costs at the model layer plus open-weight models at a fraction of the price push buyers to the cheapest adequate model, making list price the competitive variable. Frontier labs cut published API prices repeatedly, compressing revenue per token toward the cost of serving and eroding model-layer margins.
The case for no
The case against it
Every inference has a real cost in tokens and GPU cycles that does not trend toward zero with scale, so distribution becomes a survival system rather than a growth channel because undisciplined distribution scales adoption and bleeds faster with every user added. It would show first if flagship frontier list prices hold flat or rise through 2027 while labs push premium reasoning tiers.
6 voices
Model prices fall toward cost
Because routing layers and code wrappers optimize for cost, cheaper Chinese open-weight models pull volume away from OpenAI, forcing OpenAI into defensive price cuts, while Anthropic's coding-specific premium positioning shields it from the same…
Cheap Chinese open-weight models force price cuts that break OpenAI's revenue math
Cheap open-source challenger compresses frontier lab pricing
When capability is broadly available and cheap, competitive pricing pushes the gains through to buyers, as happened with jet transportation, vaccines and the PC, instead of accruing to owners as in search, social and e-commerce.
Consumer subscriptions charge fixed fees against variable compute so power users lose money, while API buyers face effectively zero switching costs and route to whatever is cheapest at the moment. Alongside that, falling software creation costs…
Open-source and low-cost competition disciplining AI pricing
1 voice
The case against it
Miqdad Jaffer
In SaaS, marginal costs approach zero once the product is built, so overpaying for acquisition was acceptable because unit economics improved at scale. In AI, every query has a price tag that often gets worse as the company grows. A free tier that…
What would settle it
The dated tests.
Each test names the public source that settles it, the rule and the day. The Call’s number, the readers’ median and the machines’ numbers stand beside it, as on the Docket.
Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 65.5%.
On the Docket Pendingth130-t04Open question
Specification pending
The machinesClaude 94% · DeepSeek 93%
Your numberSign in by email to forecast
Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 67.3%.
On the Docket Pendingth130-t05Open question
Specification pending
The machinesClaude 80% · DeepSeek 94%
Your numberSign in by email to forecast
Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 69%.
On the Docket Pendingth130-t06Open question
Specification pending
The machinesClaude 88% · DeepSeek 94%
Your numberSign in by email to forecast
C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 29.5%.
On the Docket Pendingth130-t01Open question
Specification pending
The machinesClaude 88% · DeepSeek 95%
Your numberSign in by email to forecast
C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 32%.
On the Docket Pendingth130-t02Open question
Specification pending
The machinesClaude 13% · DeepSeek 95%
Your numberSign in by email to forecast
C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 34.5%.
On the Docket Pendingth130-t03Open question
Specification pending
The machinesClaude 22% · DeepSeek 82%
Your numberSign in by email to forecast
In The Buildout
The Parts on the same question.
The stories that take up this thesis, with their own Calls and tests.
Part 12: The Model Is Not the Price
The token is cheap and getting cheaper; the labs' revenue is counted in the tens of billions, and none of it is audited on EDGAR. DeepStack reads the model's price as a four-month lease and the lab's price as a line in a filing the public has yet to read.
A thesis page states one question and both sides of it; DeepStack takes no side in words, and its view is the number it gives each test before the date. Analysis and opinion, not investment advice. Voices are paraphrased from public writing and attributed by name; DeepStack shows no score, ranking or accuracy for any person. Pre-specified means the words were approved on or before the test’s date. Specified after the date means they were written afterwards: such a test is counted and shown, and never rated. An open question has no words yet. The Call’s number is the chance, from 0 to 100, that the Call gives the yes answer to a test’s question, recorded before the test’s date; it is DeepStack’s own reading, not a market price and not advice. The readers’ mark on the same line is the readers’ median, the middle of the numbers signed-in readers gave the test, shown from 25 readers. The machines: Each model read this page and nothing else. The number that counts is the latest one each model gave before the test closed.