Multimodal Knowledge Agent Benchmark (MKA)
We show this table for reference; we do not rank on it.
Google DeepMind's benchmark of professional knowledge work where the key information sits in visual files such as architectural drawings, engineering schematics and medical scans. The agent has to deliver a finished report, spreadsheet, slide deck or marked-up drawing. Artificial Analysis runs it on a private 200-task set.
Benchmark score on MKA — November 10, 2026
We compile the MKA rows from secondary reports. Claude Fable 5.2 leads the table at 56.6%, followed by Gemini 4 Carbon (54.3%) and Claude Opus 5.5 (53.0%). We do not use these results to rank models overall.
Claude Fable 5.2
Anthropic
Gemini 4 Carbon
Claude Opus 5.5
Anthropic
19 modelsAgenticCurrentDisplay onlyUpdated November 10, 2026
Benchmark score table (19 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Fable 5.2Anthropic | 56.6% | Not reported | Closed |
| 2 | Gemini 4 CarbonGoogle | 54.3% | Not reported | Closed |
| 3 | Claude Opus 5.5Anthropic | 53.0% | Not reported | Closed |
| 4 | Claude Sonnet 5.5Anthropic | 52.0% | Not reported | Closed |
| 5 | Gemini 4 ArgonGoogle | 51.3% | Not reported | Closed |
| 6 | GPT-6 AstraOpenAI | 50.2% | Not reported | Closed |
| 7 | GPT-6.1 SolOpenAI | 47.7% | Not reported | Closed |
| 8 | Muse WonderMeta | 47.5% | Not reported | Closed |
| 9 | Kimi K3.1 ProMoonshot AI | 43.9% | Not reported | Open |
| 10 | GPT-6.1 TerraOpenAI | 43.8% | Not reported | Closed |
| 11 | Claude Haiku 5.5Anthropic | 41.0% | Not reported | Closed |
| 12 | Grok 4.7xAI | 40.1% | Not reported | Closed |
| 13 | MiMo-V2.6-ProXiaomi | 37.3% | Not reported | Open |
| 14 | GPT-6 LunaOpenAI | 36.8% | Not reported | Closed |
| 15 | GLM-5.3-FlashZ.AI | 36.2% | Not reported | Open |
| 16 | Gemini 3.8 FlashGoogle | 35.3% | Not reported | Closed |
| 17 | Mistral Large 4Mistral | 35.1% | Not reported | Open |
| 18 | MiniMax M3.1 FlashMiniMax | 33.5% | Not reported | Closed |
| 19 | Step 5 PreviewStepFun | 27.6% | Not reported | Open |
Among the reported MKA rows, Claude Fable 5.2 is first at 56.6%. The third row is 3.6 points behind. The broader top-10 range is 12.8 points, so the table still separates the published systems.
19 models have been evaluated on MKA. The benchmark falls in the Agentic category. MKA is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About MKA
Year
2026
Tasks
200 multimodal professional tasks across 10 industry verticals
Format
Climb WSR (weighted rubric satisfaction rate, %)
Difficulty
Expert multimodal knowledge work
200 private tasks built with experts from 10 verticals and 83 occupations (business, construction, education, engineering, finance, healthcare, law and government, logistics and mapping, science, software and data); each takes a professional three to eight hours. Each task has an expert rubric of about 70 weighted yes/no checks, 75% on factual accuracy and 25% on visual design. Artificial Analysis runs every model in its open-source Stirrup harness (up to 250 turns, a Linux sandbox with code execution, web search and an image viewer) and grades with an agentic judge. BenchLM shows the headline Climb WSR: the weighted share of the harder rubric checks satisfied, averaged over tasks. Google published 10 showcase tasks on Hugging Face (google/multimodal-knowledge-agent, Apache 2.0) and keeps the scored set private. Announced November 10, 2026. Independently run by Artificial Analysis under its published harness. Display-only; not yet admitted as independent-run evidence.
Freshness and provenance
Version
MKA 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does MKA measure?
Google DeepMind's benchmark of professional knowledge work where the key information sits in visual files such as architectural drawings, engineering schematics and medical scans. The agent has to deliver a finished report, spreadsheet, slide deck or marked-up drawing. Artificial Analysis runs it on a private 200-task set.
Which model scores highest on MKA?
Claude Fable 5.2 by Anthropic currently leads with a score of 56.6% on MKA.
How many models are evaluated on MKA?
19 AI models have published results on MKA in the BenchLM catalog.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.