Did he lie about the prices of Xai Grok 3?

Disputes over the AI ​​criteria – and how they informed the AI ​​laboratories – poured in public appearance.

This week, an Openai employee defendant Elon Musk’s company, Xai, the latest AI model, Groc 3, Publishing results for GROK 3. Xai, Igor Babushkin, insistently was on the right side of the company.

Between the truth is in one place.

One Writing on Xai’s BlogThe company published the final invitation from a chart showing the performance of 2025 in 2025, from the math exam. There are some experts The reliability of AIM has requested as an AI benchmark. However, the test is usually used to explore the mathematical ability of the model, 2025 and older versions.

Xai’s schedule showed two options of Grock 3, Grock 3 Reasoning Beta and GROK 3 mini-reassignment, defeated Openai’s best operation, O3-mini-highIn Aime 2025. However, Openai employees in the X, the Graph of the Article of the Article «Expencing @ 64, noted that O3-mini-height was not included in the AIME 2025.

What are the Cons @ 64, can you ask? Well, a short thing for “Consensus @ 64” and mainly a model 64, trying to answer each problem in a criterion and gets the most generated answers as the latest answers. As you can imagine, in @ 64, in @ 64, the prices of models are pricing a little and can miss a graph, as if a model can fall on another one in reality, it is not.

GROK 3 RESENTATION BETA AND GROK 3 Mini Justification in “@ 1” for 2025 points for the 2025 Grock 3 Reasoning Beta, as well as the way so far, so far O1 model Set to “Medium” computation. Still Xai Advertising Grok 3 Like “the most intelligent AI in the world.”

Babusk argued in x This Openai has published similar incorrect benchmark graphics in the past – graphics comparing the performance of their models. Another neutral party in the discussion combined a graph of more “accurate”, showing more every model performance @ 64:

Some people’s plot is cheerful, in fact, in fact, in fact, the hilarious propaganda they saw as an attack as an attack on Openai and others
(In fact, I believe that GroK looks good there, and O3-minini- * “1” “1” “” “” 1 “” “1” “” 1 “” “1” “” 1 “” “1” “” I “” “I” “” I “” I believe in the TTC Chicanery. https://t.co/djqljpcjh8 pic.twitter.com/3whefic

– Teortaxes ▶ (DeepSeek Twitter🐋iron Dust 2023 – ∞) (@TortAxestex) 20 February 2025

But as AI researcher Nathan Lambert pointed out in an articlePerhaps the most important metric remains a mystery: the calculation (and money) the value needed to achieve the best score for each model. This only shows how many AI tendencies show how much a trend is about the limitations of models and their strengths.