Method

Every model gets three ratings: intelligence, cost and value. The weights and thresholds below are read straight from the scoring config, so this page changes when the config does.

Which models are listed

The list is not picked by hand. Every day it is rebuilt from the sources: a model is listed when Artificial Analysis has an Intelligence Index for it, it is on at least one Arena board, and models.dev has an official price for it, from the company that makes it. Each model gets one row, at its highest reasoning level. Names are matched exactly across the sources, never by similarity, and the few mistakes left are corrected by hand.

A model is marked legacy, and collapsed on the home page, when the list has a newer version of it: the same name from the same company with a higher version number, as with Claude Opus 5 next to Claude Opus 5.5.

Intelligence

Each leaderboard metric is first normalised to 0–100, where 100 is the current leader on that metric. The AA Intelligence Index starts at 0. The other metrics only use a fixed range below the leader, listed in the table: Elo ratings and Agent Arena scores have no meaningful zero, and LiveBench scores sit too close together to tell models apart from 0.

n = clamp(100 × (x − floor) ÷ (frontier − floor), 0, 100)

The score is the weighted mean of the metrics a model has. A model missing from a leaderboard is scored on the rest, and a model on none of them is not rated. Each leaderboard counts once: one the AA Intelligence Index already contains, such as Terminal-Bench or FrontierCode, is not used, and one that overlaps with it in part counts only for the part that does not.

LiveBench counts only for its Reasoning, Data Analysis, Research, Browser, and Design categories. Coding and Long Agentic are left out because the AA Intelligence Index already tests the same kind of task, so LiveBench weighs 61.1% of a full leaderboard: the share of its own category weights the kept categories carry.

Cost

Prices are the official ones from models.dev, in USD per 1M tokens. Blended price = (7 × cache read + 2 × input + 1 × output) ÷ 10. That counts about 78% of input tokens as cache hits, because agents and long chats resend their context. A model without a cache-read price pays the input price for those tokens too, and cache-write surcharges are not counted. Tiers are set by the absolute blended price.

The cost index puts a price on a 0–100 log scale: $0.04 or less is 0, $12 or more is 100. It is only used to compute value.

cost_index(price) = clamp(100 × log10(price ÷ 0.04) ÷ log10(12 ÷ 0.04), 0, 100)

Value

A low price per token says little if a model burns a lot of tokens, so value looks at what a model costs to run. Artificial Analysis publishes what it paid to run its whole Intelligence Index on each model. Every model gets the same suite, so that cost already reflects how many tokens the model burns, reasoning included. To put it on the per-token price scale it is divided by typical token use: the median, over the models listed here, of run cost ÷ blended price. A model that burns a typical number of tokens ends up at its blended price, a frugal one below it and a wasteful one above. A model without a run cost keeps its blended price. Typical token use is currently 2,443 million tokens.

effective_price = run_cost ÷ typical_token_use

Value = intelligence^0.6 × max(100 − cost index of the effective price, 5)^0.4. This is a geometric mean, so neither side can be terrible: smart but very expensive, or cheap but dim, both score low.

value = intelligence^0.6 × max(100 − cost_index(effective_price), 5)^0.4

Tiers

Data sources and attribution

A scheduled job syncs external data into snapshot files. The site build only reads those files and never goes online.

Caveats

• Coding ability is measured by Arena's WebDev board and the coding part of the AA Intelligence Index. Artificial Analysis also publishes a Coding Index, but its free API doesn't include it, so it isn't used.
• Leaderboards often list one model several times, once per reasoning level. The entry used here is the level mekiki actually uses.
• The run cost comes from Artificial Analysis's own evaluation suite, at the reasoning level of the entry mapped there. Another workload or a lower reasoning level costs differently.