Thesis · Foundation Models & AI Intelligence

Will the published list price per million output tokens of the flagship frontier models from OpenAI and Anthropic fall by 50% or more from their 2026-01-01 levels by 2027-12-31?

Open-weight substitutes hosted cheaply by third-party providers will force closed labs to cut flagship API list prices by at least half.

6 dated tests · 0 resolved · next on

The tests More theses in Foundation Models & AI Intelligence

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

Open-weight competition forces price cuts

Enterprises routing agentic workloads to hosted open-weight endpoints (OpenRouter, Cloudflare Workers AI, Together) at a fraction of closed-model cost creates direct price pressure on the model layer. Observers describe open-source models closing the performance gap while remaining cheaper, pulling developer adoption away from proprietary options and compressing margins. This pressure leads a major closed lab to cut its flagship output-token list price by 50% or more from 2026-01-01 levels before 2027.

The case for no

Flagship prices hold firm

Feedback loops from reasoning models combined with reinforcement learning make frontier large language models more differentiated and harder to replicate, creating a scale-driven network effect that narrows the field of leading providers. This differentiation reduces the risk that frontier models become commodities, letting closed labs sustain flagship pricing. On this view, flagship closed-model output-token list prices from OpenAI and Anthropic stay flat or higher at end-2027 versus their 2026-01-01 levels rather than falling by 50% or more.

6 voices

Open-weight competition forces price cuts

  1. Wyndo

    The full model needs roughly 256 GB of RAM, which blocks local use on ordinary machines, but pointing Claude Code's environment variables at an OpenRouter-hosted GLM 5.2, or OpenCode at Cloudflare Workers AI, keeps the workflow identical while…

    Source

  2. Stephen Clapham

    As open source models close the performance gap with proprietary systems while remaining months behind, competition intensifies and customers can choose lower-cost alternatives, compressing margins and making the model layer unproductive for capital.

    Source

  3. Michael Spencer

    When open-source models match or exceed closed-source performance while remaining transparent and cost-effective, developers adopt them over proprietary options regardless of which company developed them.

    Source

  4. Scott Galloway

    Chinese AI models achieve similar or better performance than American LLMs at a fraction of capital expenditure. Widespread adoption of Chinese open-source models by major startups and established companies like Airbnb reduces differentiation and…

    Source

  5. Arctan Research

    The oligopoly case rested on the premise that progress required continuously larger capital outlays, which limited the field to firms able to fund them. Demonstrating that a small firm can advance training methods cheaply removes that barrier…

    Source

  6. Kevin Gee

    AOL provided a simple gateway to the Internet when the public Internet was complex. As the public Internet improved in usability and breadth, AOL's closed toll-booth became unnecessary. OpenAI is in a similar position: ChatGPT is simple and…

    Source

3 voices

Flagship prices hold firm

  1. Brown Advisory

    Feedback loops in reasoning models combined with reinforcement learning create a one-sided network effect from scale, narrowing the number of leading LLM players and reducing the risk that LLM models become commodities.

    Source

  2. James Kynge

    Data volume compounding into AI video leadership

    Source

  3. Michael Spencer

    When Mistral realizes it cannot build globally leading open-weight models despite significant resources, it pivots toward compute and infrastructure services. This move abandons its stated sovereign AI positioning and places it in direct…

    Source

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.

  1. Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 66.5%.

    On the Docket Pendingth128-t04Pre-specified

    Question
    Did the gross profit of Palo Alto Networks for the quarter ended October 31, 2026 come to at or above 66.5% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of Palo Alto Networks (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that Palo Alto Networks reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 66.5%; no if it is below 66.5%.
    If the source is late, revised or silent
    If Palo Alto Networks has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    Palo Alto Networks's gross margin was 67.6% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 73.2%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateExpects yes. Gives the yes answer 78%.

    78%

    0 · no100 · yes

    The machinesClaude 88% · DeepSeek 92%

    Claude: 88%DeepSeek: 92%
    The ladder 3 rungs on one reading, Nov 18, 2026: each asks whether it lands above its threshold, so the Call’s numbers down the rungs sketch where it expects the reading to land.

    1. Above 66.5%Call 78% · Claude 88% · DeepSeek 92%

      78%

    2. Above 68%Call 48% · Claude 94% · DeepSeek 94%

      48%

    3. Above 69.5%Call 22% · Claude 22% · DeepSeek 38%

      22%

    0 · no50100 · yes

    Your number

  2. Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 68%.

    On the Docket Pendingth128-t05Pre-specified

    Question
    Did the gross profit of Palo Alto Networks for the quarter ended October 31, 2026 come to at or above 68.0% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of Palo Alto Networks (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that Palo Alto Networks reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 68.0%; no if it is below 68.0%.
    If the source is late, revised or silent
    If Palo Alto Networks has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    Palo Alto Networks's gross margin was 67.6% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 73.2%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateEither answer fits the Call. Gives the yes answer 48%.

    48%

    0 · no100 · yes

    The machinesClaude 94% · DeepSeek 94%

    Claude: 94%DeepSeek: 94%

    Your number

  3. Palo Alto Networks, gross margin, quarter ended October 31, 2026: whether it comes in at or above 69.5%.

    On the Docket Pendingth128-t06Pre-specified

    Question
    Did the gross profit of Palo Alto Networks for the quarter ended October 31, 2026 come to at or above 69.5% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of Palo Alto Networks (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that Palo Alto Networks reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 69.5%; no if it is below 69.5%.
    If the source is late, revised or silent
    If Palo Alto Networks has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    Palo Alto Networks's gross margin was 67.6% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 73.2%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateExpects no. Gives the yes answer 22%.

    22%

    0 · no100 · yes

    The machinesClaude 22% · DeepSeek 38%

    Claude: 22%DeepSeek: 38%

    Your number

  4. C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 30.5%.

    On the Docket Pendingth128-t01Pre-specified

    Question
    Did the gross profit of C3.ai for the quarter ended October 31, 2026 come to at or above 30.5% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of C3.ai (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that C3.ai reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 30.5%; no if it is below 30.5%.
    If the source is late, revised or silent
    If C3.ai has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    C3.ai's gross margin was 31.8% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 37.6%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateExpects yes. Gives the yes answer 66%.

    66%

    0 · no100 · yes

    The machinesClaude 82% · DeepSeek 95%

    Claude: 82%DeepSeek: 95%
    The ladder 3 rungs on one reading, Dec 2, 2026: each asks whether it lands above its threshold, so the Call’s numbers down the rungs sketch where it expects the reading to land.

    1. Above 30.5%Call 66% · Claude 82% · DeepSeek 95%

      66%

    2. Above 32.3%Call 45% · Claude 27% · DeepSeek 90%

      45%

    3. Above 34%Call 24% · Claude 16% · DeepSeek 90%

      24%

    0 · no50100 · yes

    Your number

  5. C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 32.3%.

    On the Docket Pendingth128-t02Pre-specified

    Question
    Did the gross profit of C3.ai for the quarter ended October 31, 2026 come to at or above 32.3% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of C3.ai (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that C3.ai reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 32.3%; no if it is below 32.3%.
    If the source is late, revised or silent
    If C3.ai has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    C3.ai's gross margin was 31.8% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 37.6%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateExpects no. Gives the yes answer 45%.

    45%

    0 · no100 · yes

    The machinesClaude 27% · DeepSeek 90%

    Claude: 27%DeepSeek: 90%

    Your number

  6. C3.ai, gross margin, quarter ended October 31, 2026: whether it comes in at or above 34%.

    On the Docket Pendingth128-t03Pre-specified

    Question
    Did the gross profit of C3.ai for the quarter ended October 31, 2026 come to at or above 34.0% of its revenue?
    Source
    SEC EDGAR, XBRL company facts of C3.ai (Forms 10-Q and 10-K)
    Rule
    Divide the gross profit by the revenue that C3.ai reports for the quarter ended October 31, 2026 in its 10-Q or 10-K filing on SEC EDGAR. The answer is yes if the ratio is at or above 34.0%; no if it is below 34.0%.
    If the source is late, revised or silent
    If C3.ai has not filed by December 17, 2026, or a filing restates a period the rule reads, the test is inconclusive, and the Result says what was read and what was missing.
    Base rate
    C3.ai's gross margin was 31.8% for the quarter ended July 31, 2026, by its filing on SEC EDGAR.
    Outside view
    A year earlier, for the quarter ended July 31, 2025, the same line read 37.6%, so the rungs sit on the last print.
    What would change the number
    A change in product mix, a one-time charge, or a reclassified line would move the reading.

    The Call, before the dateExpects no. Gives the yes answer 24%.

    24%

    0 · no100 · yes

    The machinesClaude 16% · DeepSeek 90%

    Claude: 16%DeepSeek: 90%

    Your number

In The Buildout

The stories that take up this thesis, with their own Calls and tests.

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.