Output price of the top-ranked AA modelno data
— USD per 1M output tokens
on track >= 25 · off < 15artificial-analysis ↗
connector returned no usable reading · checked 2026-09-07
Radar · Model frontier and open weights · T2 · 2029 · THESIS
On 2028-12-31 the top-ranked model on the Artificial Analysis Intelligence Index carries a list output price of at least $20 per million tokens on its first-party API, even as trailing-tier prices keep falling.
Capacity plans assume inference revenue per token collapses toward zero. That is true for last year's capability and false, so far, for this year's. If the frontier tier holds $20 or more, the labs keep gross margin to fund training and the demand for the highest-density racks stays anchored to a few closed models. If it breaks, the buildout is being financed on a price that no longer exists.
Registered at 70% on September 8, 2026. Engine repriced 2 times; now 62%.
Dated rungs. Each is scored on its own; the thesis does not get credit for the ladder until the rungs land.
filled bar · my probabilityhollow dot · engineamber date · due, awaiting adjudication
m1 · 2027-06-30 · 80% · The top-ranked Artificial Analysis model lists at $30 or more per million output tokens.
m2 · 2027-12-31 · 75% · The top-ranked Artificial Analysis model lists at $20 or more per million output tokens.
m3 · 2028-12-31 · 70% · The top-ranked Artificial Analysis model lists at $20 or more per million output tokens.
Registered thresholds. Status is computed from the latest public reading.
Output price of the top-ranked AA modelno data
— USD per 1M output tokens
on track >= 25 · off < 15artificial-analysis ↗
connector returned no usable reading · checked 2026-09-07
Cheapest model within 5 index points of the topno data
— USD per 1M output tokens
on track >= 5 · off < 2artificial-analysis ↗
connector returned no usable reading · checked 2026-09-07
Would raise my number
Would cut it
6 personas on one model (openai/gpt-5.6-sol); other vendors were unavailable this run, so read this as one model's persona-averaged view.
57%
grid-and-power-planner · openai/gpt-5.6-sol
Using premium flagship first-party API launches—not fixed-capability inference costs—as the reference class, I assign about a 0.20 annual hazard that competition or efficiency pushes the index leader below $20. Compounded over 2.31 years, that gives roughly 0.60 survival; I trim for leaderboard discontinuation and resolution-against ambiguity. Starting at $50 provides substantial cushion, and moving-frontier models can retain scarcity pricing while older capabilities cheapen. Conversely, a 60% list-price reduction is modest relative to historical inference-cost declines, and a cheaper model could take first place. Neither leading indicator supplied a current reading. The named index and pricing field must also remain available.
64%
semiconductor-supply-chain-analyst · openai/gpt-5.6-sol
Reference class: moving-frontier flagship API list prices, rather than constant-capability inference prices. The supplied $50 baseline leaves a 60% decline before failure, while frontier leaders can preserve premiums through scarce compute, high-effort inference, and product segmentation. Against that, rapid hardware/software cost declines and competition make a sub-$20 leader plausible by late 2028. I assign roughly a 15% annual hazard that frontier pricing crosses below $20; compounded over 2.3 years this gives about 0.69 survival, reduced to 0.64 for leaderboard discontinuation, pricing-definition changes, and snapshot risk. This is below the forecaster’s 0.70.
64%
enterprise-cio · openai/gpt-5.6-sol
Reference class: successive first-party frontier API generations since commercial LLMs began, a sparse and volatile sample. Starting at $50, failure requires more than a 60% list-price cut while the measured frontier advances; constant-capability cost declines therefore overstate downward pressure. Premium prices can persist because frontier models consume more inference compute and vendors can price-discriminate, but competition, distillation, and falling serving costs create regime-change risk. I assign a 0.16 annual hazard of falling below $20; compounded over 2.31 years gives about 0.67 survival. Allowing 0.03 for index discontinuation, endpoint ambiguity, and snapshot noise yields 0.64.
62%
credit-analyst · openai/gpt-5.6-sol
My reference class is premium flagship API pricing across successive frontier-model generations over two-to-three years. Starting from $50 provides a substantial cushion: resolution requires the qualifying leader’s standard output price to fall more than 60%. Moving-frontier capability, scarcity and high inference intensity support continued premium pricing, unlike constant-capability price trends. I assign roughly a 17% annual hazard of a pricing-regime shift below $20; compounded through 2028, then allowing modest index-discontinuation and resolution-ambiguity risk, yields 0.62. Competition, cheaper near-frontier substitutes and falling compute costs make the forecaster’s 0.7 somewhat optimistic, especially with no current indicator readings.
64%
geopolitics-policy-analyst · openai/gpt-5.6-sol
Reference class: successive frontier API launches in the three-year history supplied, where top nominal output prices stayed at $50–$75 even while constant-capability inference cheapened. The current $50 price has a 60% cushion over the threshold, and the moving leaderboard tends to select costly new capability rather than fixed capability. I nevertheless use about a 17% annual hazard that competition, architectural efficiency, or a low-priced leader pushes the top qualifying model below $20; compounded over 2.3 years, plus a small discontinuation risk, this leaves 0.64. Two or more release cycles and absent indicator data justify discounting the forecaster’s 0.70.
62%
superforecaster-statistician · openai/gpt-5.6-sol
I use a 0.60 base rate for a two-to-three-year window in which the year-end capability leader among major first-party APIs retains an output price above $20. The supplied $50 starting point helps: a miss requires a fall below $20, and moving-frontier pricing is stickier than constant-capability pricing. Against that, two model generations create meaningful chances that a cheaper rival tops the index; benchmark-normalized inference declines remain directionally relevant. I model about a 0.17 annual hazard of crossing below $20, plus a small penalty for index discontinuation or resolution ambiguity. With no live indicator readings or calibration table supplied, the forecaster’s 0.70 appears too confident.