LMArena (Chatbot Arena) · Epoch Capabilities Index · prices as of 2026-10-23
LLM API rankings
There is no best LLM API, only a best one per criterion. The two quality benchmarks do not even agree: of the 43 models both rank, one moves 43 places between them. Recomputed on every rebuild from the same sourced prices as the catalogue.
Score: Epoch Capabilities Index by Epoch AI (CC BY 4.0) · « Tied with the best »: the top of the model's interval reaches the bottom of the leader's · price: cheapest offer, input + output per million tokens, discounts included.
Price against intelligence
The efficient frontier
Only 9 of 111 scored models are Pareto-optimal: for every other one, something is both cheaper and smarter. The last step of the frontier, gpt-6.1-sol → claude-opus-5-5, costs $12.00 per extra Epoch Capabilities Index point; the first, mistral-nemo → gpt-oss-20b, costs $0.004. About ×2791 more per point at the top.
Intelligence against price · up and left is better
Same sourced prices, one criterion per board. Top five of each, in the board's own order. The quality and intelligence boards only rank models released in the last 12 months on their podium.
Quality per dollar
LMArena (Chatbot Arena) Elo per $ · higher is better
Embedding models are ranked apart: the embedding board (15 models, MTEB score per dollar).
Two benchmarks
Same models, two benchmarks, two orders
43 models carry both a LMArena (Chatbot Arena) score (a human-preference vote) and an Epoch Capabilities Index score (reasoning tests), each divided by the same cheapest price. If the criterion made no difference, the two ranks would match. Instead the median model moves 28 places, 28 move 10 or more, and the widest split is llama-3.3-70b: 43rd on one, 86th on the other.
Rank gap, LMArena (Chatbot Arena) against Epoch Capabilities Index
Places a model moves when the criterion changes · colored by maker
Ranks are each model's place on its full board, so they skip: 21 models carry an Elo but no index score, 68 the reverse. The bar spans a model's two ranks on one scale. Elo source, index source.
The denominator
Hold the benchmark still, change the price
A million tokens is not a fixed amount of text: each tokenizer cuts the same passage differently. Across the 84 offers on Cheapest open-weight LLM API whose tokenizer is public, 78 change place when priced per character instead of per token, and the median line moves 3 places. mistral-nemo stays first under all 5 definitions, but its lead runs from 79.55% on French down to 37.58% on JSON. The five widest movers:
Ranks are recomputed on the same 84 offers for every column. Corpus: four fixed passages of 1792, 2131, 1625, 1563 characters (tokenizer tax). 40 of the 124 offers cannot be ranked this way: 12 models have no public tokenizer, and 21 more use one inferred from lineage. Too thin to rank: Cheapest frontier model API (2 of 11 offers), Cheapest LLM chat API (96 of 208 offers).
Know a constraint?
Narrow the field instead of ranking it
A ranking of everything is the wrong tool once you know a constraint. With a shortlist of two, head to head puts both on the same basis (7 duels). With your own hardware, self-host an LLM costs 95 open models against the card that runs them. Otherwise, cut by size or by use case; each board states its admission rule in full.
When your task needs the last few benchmark points, and only then. On the current frontier, going from gpt-6.1-sol to claude-opus-5-5 costs about $12.00 per extra intelligence point, against $0.004 at the cheap end: roughly 2791× more per point. For tasks a mid model already handles, the premium buys nothing.
What does Pareto-optimal mean for an LLM API?
A model is Pareto-optimal when no other model is both cheaper and higher-scoring. Only 9 of the 111 scored models qualify; every other model is dominated: something else offers more intelligence for less money.
Should I always pick a model on the frontier?
On this metric, yes: a dominated model gives you less score per dollar by definition. But the benchmark does not capture everything: latency, context window, modalities, licensing or strength on your specific task can justify an off-frontier pick. Use the frontier as the default, and pay off it only for a reason you can name.
