MODELS · THE PUBLIC FRONTIER
The frontier,
priced.
Every model worth deploying, plotted by measured intelligence against how many tokens you get per dollar. Up and to the right is better on both axes. Reasoning models appear at each effort level, linked into a ladder, because more thinking buys intelligence at a price. The curve bowing toward the top-right corner is the public Pareto frontier: the models nothing else beats on both axes at once. It is a clean story. It is also not your story.
THE READING
Your frontier is not this frontier.
A public benchmark averages a thousand strangers' tasks. It cannot tell you which model is cheapest at the work that is actually yours. That answer moves the curve.
-
01
The knee is where the money is
The highest-intelligence model rarely wins on value. On the public curve, a mid-priced model reaches within a few points of the leader while returning several times the tokens per dollar. On your workflows that gap widens or vanishes. We measure which.
-
02
Open weight now sets the value end
Two of the five frontier points are open weight. For tasks that do not need the last few index points, the cheapest non-dominated model is often one you can host yourself. That changes the build-versus-buy math.
-
03
Re-weighting reorders the leaders
Score the same models on your invoices, your handoffs, your edge cases, and the ranking shifts. A model that trails by three points in general can lead by ten on your domain. The public frontier is a starting hypothesis, not the verdict.
THE FIELD
Every configuration, ranked.
Twenty-one entries across nine providers, each effort tier its own row. Sorted by intelligence. Frontier rows are shaded. Every value is public and dated, so you can check us.
| Model | Lab | Index | Cost / 1M | Tokens / $ | License | Released | Frontier |
|---|
Source: Intelligence Index v4.0 (Artificial Analysis benchmarking method), public API list prices, July 2026. Blended cost is (3 x input + 1 x output) / 4 per 1M tokens, the standard enterprise-weighted ratio. For reasoning models, each effort tier is placed by its typical thinking-token load, so higher effort shows a higher cost to run and fewer effective tokens per dollar. Open-weight prices reflect the cheapest published managed-inference host and will vary by deployment. Figures are compiled for illustration of the frontier concept and should be re-verified against provider pricing before procurement.
Changelog · 2026-07-01: Claude Fable 5 restored to the field after an 18-day export-control suspension; global availability confirmed. Next scheduled refresh 2026-07-09.