What the benchmark measures
The MKA benchmark, described by AI4Business, shifts attention from the quality of textual answers to an AI agent's ability to interpret complex visual documents and produce concrete, usable deliverables, such as documents, presentations, and analyses. The assessment is based on 200 professional-type tasks. According to the source, this is the novelty: not an abstract exercise but a test oriented toward the final output that a professional could actually use.
The tool stems from a collaboration between Artificial Analysis and Google DeepMind, again according to the source. The stated goal is to go beyond traditional benchmarks, which tend to measure partial performance. The material does not specify the scoring criteria, how the document corpus was collected, or how the usability of the deliverables is judged: these are elements that remain to be verified by anyone wishing to interpret the results rigorously.
The ranking and the cost-performance ratio
In the ranking reported by AI4Business, the highest score goes to Claude Fable 5.2. Immediately behind is Gemini 4 Carbon, which however has a much lower average cost per task. The source highlights this second aspect as the economically most interesting data point: despite a performance gap described as limited, the Google model would offer a more favorable ratio between results and spending.
The comparison between the two models should be read with caution. The source does not report numerical scores, does not indicate the percentage difference between the two results, nor does it quantify the cost gap in absolute values. It is therefore not possible to establish how far ahead Claude Fable 5.2 actually is or how much more cost-effective Gemini 4 Carbon is in terms of cost per unit of performance. The indications remain qualitative and should be treated as such.
Why it matters and what is missing
The relevance of the benchmark lies in the type of capability measured. If AI agents are to enter professional workflows, the ability to read visual documents and return ready-to-use deliverables becomes an operational requirement, not an accessory. A test built on 200 tasks of this type offers a signal closer to companies' needs than tests on individual isolated skills.
However, the picture remains incomplete. The source states that frontier models are progressing but are still far from uniformity, without specifying in which areas performance is stronger and in which it is more fragile. It does not indicate the timing of the assessment, the conditions under which the tests were run, or any methodological limitations declared by the authors. Even the article's publication date, November 2026, signals that this is a snapshot destined to age rapidly in a moving market.
In summary, the MKA benchmark proposes an evaluation criterion more closely aligned with real work and highlights a trade-off between quality and cost that organizations will have to weigh. But the numbers needed to turn this indication into a procurement decision are not present in the material: the full report and an independent verification of the results will be required.
