In Artificial Analysis's evaluation of Google DeepMind's Multimodal Knowledge Agent (MKA), Gemini 4 Argon (High) scored 51% on Climb WSR, ranking 5th among the 21 models listed on the page. It led the Science vertical and tied with Gemini 4 Carbon (High) for the lead in the Image output type. These are overall and subgroup results from the same private benchmark; they do not represent all multimodal or specialized tasks.
Suitable for: Comparing Argon's weighted benchmark performance under MKA when an agent reads visually dense project files—such as architectural plans, engineering diagrams, and clinical scans—and produces professional deliverables.
Not suitable for extrapolating to: General image question answering, everyday visual understanding, success rates across all industries, or quality, latency, and cost in production.
Applicable model version: Gemini 4 Argon (High).
Test environment or client: Artificial Analysis's Stirrup harness; a Linux sandbox, code execution, internet access, web search and scraping tools, and an image-viewing tool enabled for models that support image input.
Reasoning level and parameters: High; other sampling parameters and the provider endpoint were not specified.
Artificial Analysis published its MKA evaluation on 2026-11-10 and said it developed the benchmark in partnership with Google DeepMind. Artificial Analysis implemented it as an independent evaluation; the benchmark's joint development should be distinguished from who ran the evaluation.
The methodology page says the evaluation set contains 200 private tasks across 10 industry verticals, with a rubric comprising 14,284 criteria. Each task includes a prompt, input files, and a rubric written by domain experts; inputs may include PDFs, Office documents, spreadsheets, and images. Each model ran each task once, with up to 250 turns and 8 hours per task in Stirrup. The sandbox supports code execution, web search, and scraping. Models with image input support can use an image-viewing tool; images are generally passed at their original dimensions, but may be resized to meet provider limits.
Criteria are divided into more difficult Climb items and Regression items that check basic capabilities, and are weighted from 1 to 5. GPT-6 Luna (medium) acted as the agentic judge, inspecting submitted files and final responses in a separate sandbox and assigning a binary pass/fail judgment for each criterion; the judge could not see the input files. Climb WSR is the average across tasks of the weighted pass rates for each task's Climb criteria. MKA is an independent evaluation and is not part of Artificial Analysis Intelligence Index v5.0.1.
| Observation | Result visible on the page | Scope of interpretation |
|---|---|---|
| MKA Climb WSR | 51% | The page lists 21 models; Argon ranks 5th by Climb WSR. The score is the weighted pass rate within each task, averaged across tasks; it is not a 51% rate of tasks completed successfully in full. |
| Science vertical | Leader | The article reports that Argon (High) leads in Science among the 10 verticals; it does not give the exact score for that vertical. |
| Image output type | Tied for the lead with Gemini 4 Carbon (High) | The article reports that the two models lead on Image criteria; it does not give Argon's exact score. |
On the overall leaderboard, Claude Fable 5.2 (Max) scored 57%, Gemini 4 Carbon (High) 54%, and Claude Opus 5.5 (Max) 53%; Argon's 51% was below all three. Industry and file-type leaderboards are subgroup rankings and cannot replace the overall ranking.
Artificial Analysis's article, published on 2026-11-10, says MKA uses 200 private real-world tasks across 10 industries. Tasks require handling visually complex inputs such as plans, engineering diagrams, or clinical scans, and submitting deliverables such as slide decks, spreadsheets, reports, and annotated images.
The article reports that Argon (High) leads in the Science vertical and shares the lead with Carbon (High) for the Image output type.
The linked evaluation page lists MKA Climb WSR scores: Fable 5.2 (Max) 57%, Carbon (High) 54%, Opus 5.5 (Max) 53%, Sonnet 5.5 (Max) 52%, and Argon (High) 51%. The leaderboard shows 21/21 models.
The linked methodology page adds that the rubric has 14,284 criteria, each task is run once, the limits are 250 turns and 8 hours, and GPT-6 Luna (medium) scores submissions as an agentic judge in a separate sandbox.
The 51% is the overall Climb WSR for MKA. It aggregates weighted pass rates on the more difficult criteria for each task; it should not be restated as an overall task success rate or general multimodal accuracy.
Argon's lead in Science and its tie with Carbon in Image are subgroup leaderboard results; the page does not publish exact scores for either subgroup. They do not capture performance in other verticals or output types.
Artificial Analysis developed MKA in partnership with Google DeepMind, while the methodology page says Artificial Analysis independently ran the evaluation. Disclose both the partnership and who performed the measurement: do not describe the benchmark as independent of Google, and do not treat the partnership itself as Google's endorsement of Argon's results.
The 200 tasks are private, and per-task inputs and results are not visible. The public leaderboard is not enough to independently recalculate the 51% or assess how representative it is of other task distributions.
The results apply only to the listed Argon (High) configuration and that MKA harness, judging rules, and leaderboard snapshot; they do not establish real-time product experience, production costs, or results for other model tiers.
For public verification, see the Artificial Analysis article, its linked MKA evaluation page, and the methodology page. The methodology page publishes the evaluation set size, run limits, tools, judge, and scoring definition. Because the core task set is private, external readers cannot reproduce the score task by task.
Gemini 4 Argon