FIKA-Bench (Agents): leaderboard
FIKA-Bench (2026) asks for the exact fine-grained identity of a product, species, vehicle, landmark or cultural object in an image, with human-verified evidence for every answer and items kept only when strong closed-book models failed. Multimodal agents may search the web, browse and run reverse image search within 240 seconds per example. Strict accuracy in percent from an LLM judge that accepts only answers at the required granularity; higher is better.
Metric: Overall strict answer accuracy (%) on FIKA-Bench, open-ended fine-grained recognition questions (311 images from decontaminated public fine-grained datasets and real-life photos, each answer backed by human-verified evidence URLs, kept after strong closed-book models failed); a text-only LLM judge counts only answers at the required granularity as correct, multimodal agents with search, browsing and reverse-image-search tools, 240 seconds per example; higher is better. Source: arxiv.org. Saturation forecast: Around December 2027. 2 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 19 | #264 |
| 2 | Qwen3-VL-8B-Instruct-FP8 | 13.8 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=fika-bench-agents · How It Works · Data refreshed daily, snapshot 2026-11-21.