| | --- |
| | license: cc-by-4.0 |
| | tags: |
| | - benchmarking |
| | - llm |
| | - model-evaluation |
| | - vision |
| | - ai |
| | pretty_name: https://OpenMark.ai AI Model Emotion Detection Benchmark |
| | --- |
| | |
| | # AI Model Emotion Detection Benchmark |
| |
|
| | Benchmark results from testing 11 AI models on emotion detection from movie stills, conducted on [OpenMark](https://openmark.ai) — a deterministic AI model benchmarking platform. |
| |
|
| | ## Methodology |
| |
|
| | - **Task:** Identify emotions from 4 movie stills (varying complexity) |
| | - **Models tested:** 11 (GPT-5.2, Gemini 3 Pro, Gemini 3.1 Pro, Claude Opus 4.6, Claude Sonnet 4.6, Grok 4.1 Fast, Llama 4 Maverick, Qwen 3.5, Sonar, Gemini 3 Flash, Mistral Medium) |
| | - **Runs per model:** 3 (for stability measurement) |
| | - **Scoring:** Deterministic, task-specific evaluation |
| | - **Costs:** Real API costs tracked per task |
| |
|
| | ## Key Findings |
| |
|
| | - GPT-5.2 and Gemini 3 Pro tied at 75% accuracy |
| | - Claude Opus 4.6 ($0.025/task) scored identically to Llama 4 Maverick ($0.002/task) — 12x price difference |
| | - Half the models showed ±1.000 stability variance (changed answers across runs) |
| |
|
| | ## Source |
| |
|
| | Data generated using [OpenMark](https://openmark.ai). Full analysis: [I benchmarked 10 ai models on reading human emotions](https://dev.to/openmarkai/i-benchmarked-10-ai-models-on-reading-human-emotions-3m0b) |
| |
|
