The cheapest way to reach each level of LLM capability, and how fast that cost is falling.
Built on Artificial Analysis's measured cost per task, refreshed every three hours; the methodology has the details, and the original post has the argument. Data as of …, covering … models.
The Pareto Frontier Over Time
The cheapest model at each level of the metric, and how that frontier has moved. The tabs switch between the aggregate Intelligence Index and single evaluations that map onto specific applications.
How to read this chart
The dark line is the Pareto frontier on the date set by the slider's right handle: the cheapest way to reach a given level of the selected metric among models released by then, at the prices then in effect. Dragging the right handle back hides models released after that date and moves each model to the price it had then. Dragging the left handle forward hides models released before it, so every frontier is drawn only from models released between the two handles. The gray lines are the frontier on earlier dates, labeled at their right end and marked by ticks under the slider; clicking a tick jumps to that date. Points are colored by provider, and thin lines in the same color join the reasoning levels of one model, from lowest to highest effort. Hollow points are open weights, filled are proprietary, and hovering any point names the model behind it. The cost axis is the same measured cost per task on every tab, so switching tabs only moves models vertically. Capability scores are the latest measured values; only prices are tracked over time. The current view scores every model on the current index composition and estimates its pre-recomposition costs from the model's own price history, so earlier frontiers are comparable with today's; the archive views instead show exactly the scores and measured costs in effect at the time. Models never measured under the current composition appear only in the archives. On the current view, the Cost / Time per task toggle above the figure swaps the cost axis for Artificial Analysis's measured end-to-end time per task; speeds are only recorded going forward, so positions before a model's first speed measurement use its latest measured speed.
| Cheapest model at or above | Cost per task | Score |
|---|
Cost Records by Capability Tier
The cheapest measured cost per task achieved by any released model at or above each Intelligence Index tier, by release date. Each step is a model that set a new low for its tier.
How to read this chart
Markers are colored by provider, and hollow markers are open weights models. Each model is placed at the price in effect on each date: observed prices from August 2026 onward, and recorded price events before that (see the methodology). A dashed vertical line marks a recomposition of the underlying benchmark; the running minimum resets there, since neither scores nor measured costs are comparable across it.
| Tier | First crossed | Current record | Collapse | Halving time |
|---|
Recent Frontier Advances
Each entry is a date on which a model became the cheapest way to reach some level of the Intelligence Index, through a release or a price change. Subscribe to the Atom feed to be notified of new ones.
The method behind every number on this page, its limitations, and the handling of index recompositions are described on the methodology page. The updater, the observation history, and this page's data live in the llm-frontier repository.