ZEIT
ZEIT.FINANCE

LLM IQ has topped. They will not be much smarter. But they will now be

LLM cognitive capability (raw intelligence) has plateaued and will not improve much further, but models will become progressively cheaper, faster, and more narrowly specialized.

Potential upside
+500%
Payout6×
per $1 on the 11 core bets · 1× = money back
Market odds39%
what the market priced those 11 at when this was saved
4 of 13decided0confirmed4missed
+9

Bet timeline: 13 markets, 4 did not, first verdict Sep 30, 2026, last Jul 1, 2027

  • Will the Ornn B200 Index be at least $7.00 on December 31, 2026?, Dec 31, 2026 (in 3 months), still open
  • AI bubble burst in 2026?, Jan 1, 2027 (in 3 months), still open
  • AI model scores ≥ 90% on FrontierMath Benchmark before 2027?, Dec 31, 2026, did not happen
  • Will the Ornn B200 Index hit $5.00 by December 31, 2026?, Dec 31, 2026 (in 3 months), still open
  • Will any AI model reach 1540 Overall Arena Score by September 30?, Sep 30, 2026, did not happen
  • Will any AI model reach 1600 Math Arena Score by December 31, 2026?, Dec 31, 2026 (in 3 months), still open
  • What will OpenAI's IPO valuation be?, Jul 1, 2027 (in 9 months), still open
  • Will DeepSeek be the #2 AI lab at the end of September 2026?, Sep 30, 2026, did not happen
  • Nvidia’s Market Cap end of 2026?, Dec 31, 2026 (in 3 months), still open
  • Will Perplexity's valuation hit $40B by December 31?, Jan 1, 2027 (in 3 months), still open
  • OpenAI vs Meta — higher valuation on December 31?, Jan 1, 2027 (in 3 months), still open
  • Will Moonshot be the third-best AI Agent Lab at the end of September 2026?, Sep 30, 2026, did not happen
  • U.S. enacts AI safety bill before 2027?, Dec 31, 2026 (in 3 months), still open

This worldview holds 13 bets across 4 domains. Next verdict Dec 31, 2026.

Build your own

Tracked since Aug 22 · −1.8% on paper

#9 of 38 ranked

LLM IQ has topped. They will not be much smarter. But they will now be progressively faster and cheaper and highlys specialized

Loading performance…

What has to be true

Every claim this book is built on, and what has happened to it so far.

What shows up first

  1. The next flagship model release (e.g. GPT-5.x/GPT-6) will not show a large leap in reasoning benchmarks over its predecessorMissed

    This book backs

    NO

    What that looks like: Reviewers/benchmarks call new flagship release 'incremental' not 'leap'

    • 1540Will any AI model reach … Overall Arena Score by September 30?5%now0%did not happen

    Missed

  2. Per-token/API inference cost for frontier-tier models continues to fall sharply40%today

    This book backs

    YES

    What that looks like: Major lab cuts API price per token by >50% year-over-year

    • Will the Ornn B200 Index be at least $7.00 on December 31, 2026?No92%now27%
    • Will the Ornn B200 Index hit $5.00 by December 31, 2026?Yes77%now53%

    Markets price this at 40% today

  3. Labs emphasize cheap/fast small-model variants (mini, flash, turbo) over raw capability flagships

    This book backs

    YES

    What that looks like: New 'mini'/'flash'-class model release optimized for cost/latency, not benchmark scores

    No bet in this book covers this claim.

What confirms it

  1. No LLM will pass a rigorous general-intelligence benchmark (e.g. ARC-AGI-2) at a human-competitive level within the period89%today

    This book backs

    NO

    What that looks like: ARC-AGI-2 leaderboard remains far below human-level score

    • AI model scores ≥ 90% on FrontierMath Benchmark before 2027?No21%now0%did not happen
    • Will any AI model reach 1600 Math Arena Score by December 31, 2026?No72%now89%

    Markets price this at 89% today

  2. China narrows the AI gap with the US via cheap, efficient, specialized models rather than needing frontier-IQ breakthroughsMissed

    This book backs

    YES

    What that looks like: A Chinese lab (e.g. DeepSeek) tops cost-efficiency or open-weight leaderboards

    • Will Moonshot be the third-best AI Agent Lab at the end of September 2026?Yes48%now0%did not happen
    • Will DeepSeek be the #2 AI lab at the end of September 2026?Yes1%now0%did not happen

    Missed

  3. AI-driven job displacement continues in narrow, cost-sensitive tasks (support, coding assist) even without AGI-level intelligence

    This book backs

    YES

    What that looks like: Company cites AI automation for layoffs in a cost-sensitive function

    No bet in this book covers this claim.

What pays it off

  1. No major lab formally declares AGI achieved

    This book backs

    NO

    What that looks like: No OpenAI/DeepMind/Anthropic announcement of AGI attainment

    No bet in this book covers this claim.

  2. Specialized/narrow models (coding agents, math solvers) proliferate and outperform generalist frontier models in their niche

    This book backs

    YES

    What that looks like: A narrow specialized model beats GPT/Claude flagship on a domain-specific leaderboard

    No bet in this book covers this claim.

  3. AI infrastructure capital expenditure growth (data centers, chips) decelerates as capability gains fail to justify spend

    This book backs

    YES

    What that looks like: Nvidia or hyperscaler guidance shows AI capex growth slowing quarter-over-quarter

    No bet in this book covers this claim.

  4. AI-sector equity valuations undergo a correction as 'AI bubble' narrative takes hold19%today

    This book backs

    YES

    What that looks like: AI-heavy index or basket falls >20% from highs on growth-disappointment narrative

    • Less than $1TWhat will OpenAI's IPO valuation be?9%now22%
    • Less than $4.00TNvidia’s Market Cap end of 2026?13%now5%
    • +3 more bets

    Markets price this at 19% today

  5. Existential/AGI-risk regulation urgency fades; policy focus shifts to deployment, jobs, and cost rather than superintelligence risk95%today

    This book backs

    YES

    What that looks like: Major AI bill/executive order framed around jobs/deployment rather than existential risk

    • U.S. enacts AI safety bill before 2027?No90%now95%

    Markets price this at 95% today

  6. AI-compute-narrative crypto tokens underperform as exponential-capability story loses steam

    This book backs

    NO

    What that looks like: AI-linked crypto tokens underperform Bitcoin over the period

    No bet in this book covers this claim.

If only some of it happens

Drag to see what a partial hit pays.

If 7 of these 13 outcomes occur →≈1.9×per $1
Stake back at 10 of 13

Worst case 0.4× · typically 1.9× · best case 8.9× per $1.

Pick which ones
BetWhat it adds per $1

What brings the money back

Money back · 1 of 13this bet landing returns the full stake.

Markets priced these odds at around 13%. : around 27%.

Doubled · 1 of 13this bet landing returns 2× the stake.

Markets priced these odds at around 9%. : around 23%.

Any 10 of the 13 bets landing returns the stake, whichever 10 they are. The bets named above are the fastest route, not the only one.

Where the money sits

13 bets, ranked by how much of the stake each one holds.

+500%13 bets

U.S. enacts AI safety bill before 2027?

19%1.1×

Less than $4.00T

Nvidia’s Market Cap end of 2026?

19%7.7×

Less than $1T

What will OpenAI's IPO valuation be?

19%11.7×

AI bubble burst in 2026?

19%8.1×
Below a flat split

Will any AI model reach 1600 Math Arena Score by December 31, 2026?

3.0%1.4×

Will Perplexity's valuation hit $40B by December 31?

3.0%1.1×

OpenAI vs Meta — higher valuation on December 31?

3.0%1.2×

Will the Ornn B200 Index be at least $7.00 on December 31, 2026?

3.0%1.1×

Will the Ornn B200 Index hit $5.00 by December 31, 2026?

3.0%1.3×

Will Moonshot be the third-best AI Agent Lab at the end of September 2026?

3.0%2.1×

AI model scores ≥ 90% on FrontierMath Benchmark before 2027?

3.0%4.9×

Will DeepSeek be the #2 AI lab at the end of September 2026?

2.0%154×

1540

1.1%22.2×

Bar length is share of the stake; the longest is 19%. The second rule marks an equal split.

All 13 bets, and how each is sized.

Think something different? Price it.

Build your own