# The Multimodal Knowledge Agent (MKA) benchmark

MKA tests whether an AI agent can read visually complex work files, such as drawings, schematics and scans, and produce the report, spreadsheet or deck a professional would. Google DeepMind built it with Artificial Analysis and published it on 10 November 2026. This page explains the method and lists the current scores, each line linked to the page it comes from.

## What MKA measures

- Google describes MKA as "a new benchmark built to evaluate how agents reason across visually complex files to solve real-world knowledge work tasks." (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- Its reason for building it: "Most knowledge work benchmarks can be solved with text alone, without ever needing to look at an image or page layout." (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- Artificial Analysis says it "tests whether agents can read visually dense project files, such as architectural drawings, engineering schematics and clinical scans, and produce the deliverables a professional would." (source: https://artificialanalysis.ai/evaluations/mka)
- Each task gives the agent a workspace of input files and asks for deliverables such as a report, a spreadsheet, a slide deck or a marked-up drawing. (source: https://artificialanalysis.ai/evaluations/mka)

## Who made it

- Google DeepMind developed MKA in partnership with Artificial Analysis and announced it on 10 November 2026. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- Domain experts across 10 verticals and 83 occupations, from architects and biologists to underwriters and data scientists, created the 200-task evaluation set. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- Artificial Analysis runs the evaluation and keeps the 200 tasks private, so they cannot leak into training data or be found by a model through search. A separate set of 10 public tasks is on Hugging Face. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)

## How it is run and scored

- 200 private tasks across 10 industry verticals. Each task has a prompt, a set of input files (PDFs, Office documents, spreadsheets and images) and a rubric written by domain experts. Each model makes one run per task. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- Each task reflects a multi-step project that takes a seasoned professional three to eight hours. Its rubric holds roughly 70 yes/no checks, weighted 75% on factual accuracy and 25% on visual design. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- Across all tasks the rubrics hold 14,284 criteria, each with a weight from 1 to 5. Each criterion is labelled Climb (the harder ones) or Regression (baseline competence). (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- Every model runs in the open-source Stirrup harness with up to 250 turns and 8 hours per task, in a Linux sandbox with the input files and internet access. Its tools are code execution, web search, page fetch, and an image viewer that is turned on only for models that accept images. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- A GPT-6 Luna (medium) agentic judge grades each run in a separate sandbox with no network access. It sees the task prompt, the submitted files, the final response and each criterion, but not the input files, and returns one yes or no per criterion. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- When domain experts graded model outputs themselves, they agreed with the autorater 96% of the time. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)
- A run that submits no files and no text response scores 0, and so does a run that errors. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- A task's weighted satisfaction rate (WSR) is the weight of its satisfied criteria divided by the weight of all its criteria. The headline score, Climb WSR, is the mean over tasks of the WSR for Climb criteria. Regression WSR is the same mean for Regression criteria. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- On the same tasks, the models Google tested used between 30 and 130 turns on average to finish a task. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)

## Current MKA leaderboard

21 models, read on artificialanalysis.ai on 2026-11-11 from https://artificialanalysis.ai/evaluations/mka. Climb WSR in percent, higher is better. looot did not run the benchmark; the scores are AA's as published. Each model name links to its Artificial Analysis page.

| Rank | Model | MKA Climb WSR | Source |
|---|---|---|---|
| 1 | [Claude Fable 5.2 (Max, Default Fallback)](https://artificialanalysis.ai/models/claude-fable-5-2) | 56.6% | https://artificialanalysis.ai/evaluations/mka |
| 2 | [Gemini 4 Carbon (High)](https://artificialanalysis.ai/models/gemini-4-carbon) | 54.3% | https://artificialanalysis.ai/evaluations/mka |
| 3 | [Claude Opus 5.5 (Max, Default Fallback)](https://artificialanalysis.ai/models/claude-opus-5-5) | 53.0% | https://artificialanalysis.ai/evaluations/mka |
| 4 | [Claude Sonnet 5.5 (Max, Default Fallback)](https://artificialanalysis.ai/models/claude-sonnet-5-5) | 52.0% | https://artificialanalysis.ai/evaluations/mka |
| 5 | [Gemini 4 Argon (High)](https://artificialanalysis.ai/models/gemini-4-argon) | 51.3% | https://artificialanalysis.ai/evaluations/mka |
| 6 | [GPT-6 Astra (Max)](https://artificialanalysis.ai/models/gpt-6-astra) | 50.2% | https://artificialanalysis.ai/evaluations/mka |
| 7 | [GPT-6.1 Sol (Max)](https://artificialanalysis.ai/models/gpt-6-1-sol) | 47.7% | https://artificialanalysis.ai/evaluations/mka |
| 8 | [Muse Wonder (Max)](https://artificialanalysis.ai/models/muse-wonder) | 47.5% | https://artificialanalysis.ai/evaluations/mka |
| 9 | [Kimi K3.1 Pro](https://artificialanalysis.ai/models/kimi-k3-1-pro) | 43.9% | https://artificialanalysis.ai/evaluations/mka |
| 10 | [GPT-6.1 Terra (Max)](https://artificialanalysis.ai/models/gpt-6-1-terra) | 43.8% | https://artificialanalysis.ai/evaluations/mka |
| 11 | [Claude Haiku 5.5 (Max)](https://artificialanalysis.ai/models/claude-haiku-5-5) | 41.0% | https://artificialanalysis.ai/evaluations/mka |
| 12 | [Grok 4.7 (Xhigh)](https://artificialanalysis.ai/models/grok-4-7) | 40.1% | https://artificialanalysis.ai/evaluations/mka |
| 13 | [DeepSeek V4.1 Flash 1017 (Max)](https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017) | 38.5% | https://artificialanalysis.ai/evaluations/mka |
| 14 | [Qwen3.8 Max (0902)](https://artificialanalysis.ai/models/qwen3-8-max) | 37.4% | https://artificialanalysis.ai/evaluations/mka |
| 15 | [MiMo-V2.6-Pro](https://artificialanalysis.ai/models/mimo-v2-6-pro) | 37.3% | https://artificialanalysis.ai/evaluations/mka |
| 16 | [GPT-6 Luna (Max)](https://artificialanalysis.ai/models/gpt-6-luna) | 36.8% | https://artificialanalysis.ai/evaluations/mka |
| 17 | [GLM-5.3-Flash](https://artificialanalysis.ai/models/glm-5-3-flash) | 36.2% | https://artificialanalysis.ai/evaluations/mka |
| 18 | [Gemini 3.8 Flash (High)](https://artificialanalysis.ai/models/gemini-3-8-flash) | 35.3% | https://artificialanalysis.ai/evaluations/mka |
| 19 | [Mistral Large 4 (High)](https://artificialanalysis.ai/models/mistral-large-4) | 35.1% | https://artificialanalysis.ai/evaluations/mka |
| 20 | [MiniMax-M3.1-Flash](https://artificialanalysis.ai/models/minimax-m3-1-flash) | 33.5% | https://artificialanalysis.ai/evaluations/mka |
| 21 | [Step 5 Preview](https://artificialanalysis.ai/models/step-5) | 27.6% | https://artificialanalysis.ai/evaluations/mka |

Google says Gemini 4 Carbon "achieves #2 in quality among the models we tested, and completes tasks at 6x lower cost on average compared to the top scoring model." (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)

## What these numbers do not say

- Each score comes from one run per task, made by Artificial Analysis. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- MKA is a standalone evaluation. It is not part of the Artificial Analysis Intelligence Index. (source: https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka)
- The judge is itself a model (GPT-6 Luna, medium). Google reports 96% agreement with domain experts, not 100%. (source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/)

## Sources

- [Google: Introducing the Multimodal Knowledge Agent benchmark (10 November 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/multimodal-knowledge-agent-benchmark/), read on 2026-11-11
- [Artificial Analysis: Multimodal Knowledge Agent (MKA) Benchmark](https://artificialanalysis.ai/evaluations/mka), read on 2026-11-11
- [Artificial Analysis: MKA methodology](https://artificialanalysis.ai/methodology/intelligence-benchmarking#mka), read on 2026-11-11