Claude Fable 5.2 leads the MKA benchmark with 57%, but one task costs $31.97

Most agentic benchmarks still test how well a model writes code or navigates websites. Real office work looks different: an engineer receives an architectural blueprint, an analyst receives a scanned document, and both are expected to deliver a finished file, not a chat answer. That is what MKA (Multimodal Knowledge Agent) tests. It is a new Google DeepMind benchmark that Artificial Analysis launched exclusively on its platform.

What the benchmark tests

The benchmark has 200 closed tasks across 10 industries. The agent works with dense visual materials: architectural blueprints, engineering diagrams and clinical scans. It must produce one of six output file types: a presentation, a spreadsheet, a PDF, a Word document, a text file or an annotated image.

The main difference from many other tests is that the image here is not just extra context. The task cannot be solved without a careful analysis of the visual files. If you build an agent for a construction company or a clinic, this test is closer to your reality than yet another set of coding tasks.

Each task is scored against expert rubrics, with about 70 weighted criteria on average. The criteria fall into two groups. "Regression" covers baseline checks that a model is required to pass. "Climb" covers hard checks that separate strong models from mediocre ones. The main metric, Climb WSR, shows the weighted share of hard criteria met.

The benchmark runs in Stirrup, an open agent framework. The agent has code execution, web search, page fetching and image viewing, with up to 250 turns. An agentic judge based on GPT-6 Luna (medium) checks the results.

Who leads and what it costs

Claude Fable 5.2 (max) is in first place with a Climb WSR of 57%. Next come Gemini 4 Carbon (high) at 54% and Claude Opus 5.5 (max) at 53%. Looking at the numbers soberly, the best model meets just over half of the hard criteria. It is still far from a reliable replacement for a specialist doing this kind of work.

Cost is the most telling metric. Fable 5.2 spends $31.97 per task on average. That is more than 5 times the price of Gemini 4 Carbon ($6.29) for roughly two extra points of quality. It is also about 29 times the price of GPT-6.1 Sol ($1.11). All four OpenAI models tested (GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna) sit on the price-quality Pareto frontier below Carbon. Claude Haiku 5.5 and MiMo-V2.6-Pro are there too.

Here, the Pareto frontier means something simple: you cannot swap these models for another one that is both cheaper and better.

There is no universal winner

The picture varies by industry:

• Claude Fable 5.2 leads in five verticals: Software & Data, Engineering, Logistics & Mapping, Business and Finance (it shares the last one with Claude Sonnet 5.5).
• Claude Opus 5.5 is ahead in education and construction.
• Gemini 4 Argon wins in science, and Gemini 4 Carbon wins in healthcare.
• Claude Sonnet 5.5 leads by a small margin in law and the public sector.

By output file type, the spread is smaller than by industry. Fable 5.2 is ahead on PDFs, Word documents, spreadsheets (by a minimal margin) and presentations (together with Opus 5.5). Gemini 4 Argon and Carbon are best with images, and DeepSeek V4.1 Flash 1017 is best with text files. The hardest formats are Word and text files: no model scored above 52%.

What to do with this

The first takeaway for product managers: do not choose a model by the overall ranking. If your use case is medicine or legal documents, the leader may be a different model, and you should test models on your own vertical.

The second takeaway: count the cost per task, not per token. A model at $1.11 per task may make more sense than one at $31.97 if the quality gap is only a few points and a human reviews the result anyway.

There are also limitations. The tasks are closed, so you cannot reproduce the results yourself. A model also does the scoring, even if it works from expert rubrics. The judge is GPT-6 Luna, and OpenAI models appear in the same table, so bias cannot be fully ruled out. One more caveat: the benchmark measures the quality of the finished files, not safety or speed. It is a useful reference point for choosing a model for visually heavy office work, but it is better to make the final decision after a run on your own files.