The model that came first costs 178 times the cheapest one
When Google DeepMind and Artificial Analysis launched the Multimodal Knowledge Agent benchmark on 10 November, Google’s own write-up put Gemini 4 Carbon second and did not name the model above it. The public leaderboard now does. Artificial Analysis’s MKA table is led by Anthropic’s Claude Fable 5.2 in its Max configuration, at a Climb weighted satisfaction rate of 57%.
Gemini 4 Carbon at its high setting is second on 54%, and Claude Opus 5.5 in Max configuration is third on 53%. Twenty-two of the twenty-three models evaluated appear in the chart.

Three points, and a bill five times larger
The ranking is close. The cost column is not.
Artificial Analysis puts Claude Fable 5.2 at an average of $31.97 per task. Gemini 4 Carbon, three points behind, runs at $6.29 per task — roughly a fifth of the price for a result within the margin most leaderboards would call noise. GPT-6.1 Sol costs $1.11 per task, which Artificial Analysis notes makes the top model about 29 times more expensive.
At the other end of the table, GPT-6 Luna in Max configuration completes a task for $0.18, with Claude Haiku 5.5 and MiMo-V2.6-Pro both at $0.32. The leader costs about 178 times what the cheapest entry does.
That spread is the finding. On a benchmark built to approximate paid knowledge work, the best score available is purchasable, but the premium over a result three points lower is not a rounding difference in a procurement budget.

What the benchmark actually asks
MKA is a private set of 200 tasks across 10 industry verticals. It gives an agent visually dense professional files — architectural drawings, engineering schematics, clinical scans — and asks for a deliverable: a report, a spreadsheet, a slide deck, a marked-up drawing.
Submissions are graded against expert rubrics by an agentic autorater, which Artificial Analysis says is GPT-6 Luna at its medium setting. Tasks run in a harness Artificial Analysis calls Stirrup. The headline Climb weighted satisfaction rate is the weighted share of the harder rubric criteria a task satisfies, averaged across tasks.
Two caveats travel with the numbers. The benchmark was co-designed by Google DeepMind, one of the vendors whose models it ranks, and the grader is a model from OpenAI, another. Neither fact invalidates the table, but both belong next to it.
Where Google’s newest model sits
Google’s Gemini 4 Hydrogen announcement on 12 November reports the new model at 53.4% on MKA, which would place it below both Carbon and the Anthropic entries at the top. Google published no cost figure for Hydrogen on this benchmark.
The thing to watch is whether the cost column starts moving faster than the score column. A three-point gap that costs five times more is a question about pricing, not capability, and pricing is the easier of the two to change.