The most expensive AI model we tested is not the best

You expect a pricier model to see better. In our tests that is often false.

The most expensive model on the index is Claude Fable 5.1 from Anthropic. Answering 1,000 questions about pictures costs about $88 at Anthropic’s list price. Its overall score is 69.6 out of 100, which puts it ninth of the 36 models. Eight models score higher, and every one of them costs less. Gemini 4 Argon from Google scores 77.2 for $6.52. GPT-6.1 Sol from OpenAI scores 81.3 for $7.94. That is 13 times and 11 times less money, for a higher score.

The second most expensive, Kimi K3.1 Pro from Moonshot, costs $71.79 and scores 70.6. It too is beaten by GPT-6.1 Sol and Gemini 4 Argon at about a tenth of the price.

Anthropic’s Claude Fable 5.2 shows the same thing from the other side. It costs $53.10, 40% less than Fable 5.1, and scores 80.7, 11.1 points more. GPT-6.1 Sol is level with it at about a seventh of the price.

Inside one family, more money stops helping

Anthropic’s five models show it plainly:

From Haiku to Sonnet you pay 11 times more and gain 7.5 points. From Sonnet to Opus you pay 2.3 times more and gain 5.5. From Opus to Fable 5.2 you pay 2.0 times more and gain 4.2. From Fable 5.2 to Fable 5.1 you pay 1.7 times more and lose 11.1. Against Opus, Fable 5.1 costs 3.4 times more and scores 6.9 less. Fable 5.1 is no better than Sonnet: their margins of error overlap. Fable 5.1 costs 7.6 times as much as Sonnet.

OpenAI’s family flattens in a different way. GPT-6 Luna ($0.74, 60.7) to GPT-6.1 Sol ($7.94, 81.3) is a jump of 20.6 points for 10.7 times the price. From Sol to GPT-6 Astra ($38.32, 83.7) you pay 4.8 times more for 2.4 points.

Google shows that newer can beat bigger. Gemini 3.8 Flash scores 61.9 at $12.80. Gemini 4 Argon scores 77.2 at $6.52, 15.3 points higher and half the cost. It also matches Claude Opus 5.5 (76.5) at about a quarter of Opus’s $26.16. Google’s open model Gemma 4 31B sits low on its line: 38.8 for $1.94.

Same price, very different results

Qwen3.8 Max from Alibaba and Grok 4.7 from xAI both cost $27.40 per 1,000 answers. Qwen3.8 Max scores 70.6, level with Kimi K3.1 Pro, in the group that shares sixth to ninth place. Grok 4.7 scores 54.9, in the group that shares 10th to 18th place. Both were tested on a smaller sample, so their margins are wide. Still, their ranges do not overlap: 66.1 to 75.2 for Qwen3.8 Max, 48.1 to 64.1 for Grok 4.7.

How we price it

All prices are the standard list price of the tokens used for 1,000 answers. For the Claude models we counted the tokens each answer used (the picture, the question and the answer) and priced them at Anthropic’s published list price. The price includes the tokens a model spends thinking before it answers. (An earlier version of this page used the cost reported by Anthropic’s command-line tool, which bills the whole question at a higher caching rate; that put the Claude prices 40 to 65% too high. The conclusion did not change.)

What this does not say

We measured eyesight, not everything. Fable 5.1 may be better at writing, coding or long reasoning. For looking at pictures and answering a question, we found no reason to pay its price. Fable 5.1, Kimi and several other models were run on an evenly spread sample of each test, so their margins of error are wider than those of models run on every question. Even so, Opus and Sol are clearly above Fable 5.1, and Argon is too.

How we tested

Every model got the same questions in the same format. Each question has one exact answer built into the picture, so answers are scored automatically. The score is corrected for guessing: 0 is no better than a guess, 100 is always right. Cost is the cost of 1,000 such answers. The pictures and questions are never published, so models cannot have learned them from the web.

Tested between 22 October and 10 November 2026. Each model’s own dates are on the leaderboard. See also the method.