56 Evaluations Reshuffle AI Model Leaderboard: Opus 5.5 Tops Coding, Gemini 4 Argon Tops Office Work

AIHOT's model leaderboard expanded its public evaluation set from 39 to 56 and recalculated ELO. Models that have not officially launched were excluded, such as Gemini 4 Carbon, which has 25 evaluations and a score of 75.5 but is not ranked for now. The updated overall leaderboard shows 30 models; the 56 evaluations come from 16 organizations. The page was updated at 2026-11-02T02:05:00.000Z.

Claude Opus 5.5 ranks first overall with 74.3 points and 100% evaluation coverage. It also ranks first in coding. GPT-6 Astra leads research and knowledge with 71.9 points. Gemini 4 Argon leads office work with 71.6 points. In the tweet, Khazix said this ranking matches his understanding of all the models.

The algorithm only considers scores actually measured by third-party evaluation organizations; vendor-published scores are not counted. After conversion, each result is weighted by domain: coding 38%, research 19%, professional office work 19%, knowledge Q&A 14%, and human blind selection 10%. A score of 50 is the average level of the reference models, and every 15 points is roughly one standard deviation. When two models differ by more than 3 points, their order basically will not change. Price does not affect ranking, and missing evaluations are neither filled in nor penalized.

In the comments, some people agreed with the ranking, while others focused on the move to exclude unreleased models. One user said it was right to exclude Carbon, which has not officially launched; otherwise the top spot would always be a model nobody can use. Others felt category-specific rankings are more useful than the overall leaderboard: choose the coding leader for writing code and the office-work leader for making spreadsheets; when doing real work, nobody uses just one model. The overall No. 1 is more like a default value for people who don't want to choose.

Gemini 4 Argon taking first place in office work surprised some people, while others found it reasonable. The latter's reasoning is practical: Google has its own full suite of Docs and spreadsheet products, giving the model plenty of opportunities for daily practice. However, some users reminded that the leaderboard can only serve as an initial screening; before deployment, run your own real tasks through it and factor in price and speed.

Some also wanted a before-and-after ranking comparison table for the newly added evaluations. The models didn't change, but the test paper did; which rankings remain stable is more informative than just looking at No. 1. Another user asked why Grok 4.7 was missing. Khazix's reply was blunt: because it's too weak, so it doesn't make the first screen.

After this expansion, the top few places didn't change much. One user said people argue fiercely online, but in actual use the reputation is quite stable. Another reality is that research and knowledge point to GPT-6 Astra, while coding points to Claude Opus 5.5; to truly cover these scenarios, you probably can't avoid subscribing to both. The leaderboard and evaluation details for each model are at https://aihot.news/leaderboard; you can click through and see for yourself.