In one MKA slide-task example published by Artificial Analysis, Gemini 4 Carbon (high) took about 1 hour and 76 turns, cost about $8, and scored 72%. Claude Fable 5.2 (max) took about 3 hours and 224 turns, cost about $68, and scored 87%. This is a record of one task, not an overall model ranking.
What this can help assess: The publicly reported time, number of turns, cost, number of image views, and slide score for these two models on this architectural drawing review and slide-delivery task.
What should not be generalized: These results cannot establish which model is better overall on other MKA tasks, other output types, or general knowledge work. They also do not establish a general relationship between time, number of tool-call turns, image views, and score.
Applicable model versions: Gemini 4 Carbon (high); Claude Fable 5.2 (max).
Test environment or client: An MKA example posted by Artificial Analysis on X. The execution endpoint, tool configuration, and hardware environment are not specified.
Reasoning tiers and parameters: Carbon used high and Fable 5.2 used max. Other parameters are not specified.
The original post says both models completed the same task. An earlier post in the thread explains that this MKA example comes from a public 10-task dataset. An agent inspects architectural floor plans and sections, checks for inconsistencies in doors and windows, and organizes its findings into slides.
The same thread says MKA covers 10 verticals and six deliverable file types. Its scoring criteria are divided into Climb (harder criteria that frontier models still often fail) and Regression (benchmark capabilities). The original post does not specify which category this architectural slide task falls into, and does not disclose the task's scoring rubric, review process, run configuration, or number of repetitions.
| Model and tier | Turns to completion | Time | Cost | Slide score | Image views |
|---|---|---|---|---|---|
| Gemini 4 Carbon (high) | 76 | About 1 hour | About $8 | 72% | 118 |
| Claude Fable 5.2 (max) | 224 | About 3 hours | About $68 | 87% | 203 |
These are approximate figures or task-level reported values from the original post. It does not provide billing details for these metrics, define what counts as a turn, or explain how image views were counted.
The original post says Carbon and Fable 5.2 performed “the same task” and lists the number of turns, approximate time, approximate cost, slide score, and number of image views for each.
Nearby posts in the thread describe the task as reviewing architectural floor plans and sections for door and window mismatches, then reporting the findings in slides.
The thread says the public dataset contains 10 tasks and MKA covers 10 verticals. It does not show the full task prompt, input files, or either completed slide deck.
The thread mentions the Climb and Regression scoring categories, but does not explain how the 72% and 87% in the original post were calculated or provide the task's scoring rubric.
This record compares only one slide-delivery task in MKA. In the figures reported by the original post, Carbon used fewer turns, took less time, and cost less, while Fable 5.2 earned the higher slide score. A single example cannot represent either model's performance across the full MKA suite, all 10 verticals, or other workloads.
This is a task-level case and uses different data from the composite Intelligence Index score and cost statistics in Evaluation 02. The 72% and 87% here are not Intelligence Index scores and cannot be compared directly with them.
Review the original post and adjacent thread posts to check the task description and reported figures. Reproducing the comparison would also require the original task and files, runtime model and tool configuration, completed slide decks, scoring rubric, and cost calculation method. These details are not provided in the posts currently visible.
Gemini 4 Carbon