Gemma 4 31B vs GPT-5.6 Luna
Compare Gemma 4 31B and GPT-5.6 Luna side-by-side. See how these vision models stack up in Image Captioning, OCR, Open Prompt, Object Detection, and Classification.
Compare Gemma 4 31B vs GPT-5.6 Luna live
Run the same image across every model that supports a task and compare their outputs side-by-side.
Detect and compare bounding boxes across models on the same image.
Upload an image
Drag and drop an image here, or click to browse
Models in this comparison
Gemma 4 31B vs GPT-5.6 Luna on Vision Evals
GPT-5.6 Luna scores higher on 4 of the six Vision Evals tasks.
The widest gap is Counting, where GPT-5.6 Luna leads 67.1% to 51.4%.
Overall, Gemma 4 31B averages 67.0% (#38 of 61) against 73.8% (#24 of 61) for GPT-5.6 Luna.
GPT-5.6 Luna is both cheaper ($0.0010 vs $0.0015 per sample) and faster (7.4s vs 34.4s per sample).
Gemma 4 31B vs GPT-5.6 Luna Comparison Table
Evals updated October 8, 2026Pricing updated October 8, 2026
| Property | Gemma 4 31B | GPT-5.6 Luna |
|---|---|---|
| Organization | OpenAI | |
| Category | open | closed |
| Modality | multimodal | multimodal |
| Release Date | Apr 2026 | Jul 2026 |
| Context Window | 256K | 1.5M |
| Parameters | 31B | Unknown |
| License | Apache 2.0 | Proprietary |
| Pricing per 1M tokens | ||
| Input $/1M | $0.090 | $0.200 |
| Output $/1M | $0.340 | $1.20 |
| Vision Tasks | ||
| Captioning | Demo | Demo |
| Chart Question Answering | Supported | Supported |
| Classification | Demo | Demo |
| Document Question Answering | Supported | Supported |
| Image Tagging | Supported | Supported |
| Multi-Label Classification | Supported | Supported |
| Object Detection | Demo | Demo |
| OCR | Demo | Demo |
| Vision Language | Supported | Supported |
| Visual Question Answering | Demo | Demo |
| Model Features | ||
| Foundation Vision | Supported | Supported |
| LLMs with Vision Capabilities | Supported | Supported |
| Multimodal Vision | Supported | Supported |
Vision Evalsground-truth scores across 6 vision tasks, pooled at low effort | ||
| Overall | 67.0% | 73.8% |
| Quantizationsself-hosted | ||
| Avg cost / sample | $0.0015 | $0.0010 |
| Avg speed / sample | 34.36s | 7.38s |
| By task | ||
| Object Detection (low) | 47.5% ±0.6, Mean of 3 runs, range 46.8 to 48.0 | 61.0% ±1.2, Mean of 3 runs, range 59.9 to 62.2 |
| Object Detection (high) | – | 62.3% ±1.2, Mean of 3 runs, range 61.4 to 63.8 |
| Counting (low) | 51.4% ±2.7, Mean of 3 runs, range 48.6 to 54.0 | 67.1% ±1.4, Mean of 3 runs, range 66.2 to 68.9 |
| Counting (high) | – | 70.7% ±3.4, Mean of 3 runs, range 66.2 to 73.0 |
| Identification (low) | 79.2% ±1.6, Mean of 3 runs, range 78.1 to 81.3 | 83.3% ±1.6, Mean of 3 runs, range 81.3 to 84.4 |
| Identification (high) | – | 84.4% ±6.3, Mean of 3 runs, range 78.1 to 90.6 |
| OCR (low) | 90.8% ±0.6, Mean of 3 runs, range 90.2 to 91.5 | 90.7% ±1.8, Mean of 3 runs, range 88.4 to 92.0 |
| OCR (high) | – | 91.5% ±0.3, Mean of 3 runs, range 91.2 to 91.7 |
| Data Extraction (low) | 80.4% ±1.0, Mean of 3 runs, range 79.4 to 81.4 | 80.4% ±2.1, Mean of 3 runs, range 78.3 to 82.5 |
| Data Extraction (high) | – | 81.8% ±0.5, Mean of 3 runs, range 81.4 to 82.5 |
| Reasoning (low) | 52.8% ±1.3, Mean of 3 runs, range 51.7 to 54.3 | 60.5% ±5.0, Mean of 3 runs, range 55.0 to 64.9 |
| Reasoning (high) | – | 65.6% ±3.6, Mean of 3 runs, range 60.9 to 68.2 |
Gemma 4 31B vs GPT-5.6 Luna: Overview
Gemma 4 31B is the largest dense model in Google's Gemma 4 family, built from the same research as Gemini 3 and released as open weights under the Apache 2.0 license. It supports a 256K token context window with text and image input, configurable thinking mode for step-by-step reasoning, and multilingual support across 140+ languages. The unquantized model fits on a single 80GB GPU.
For vision tasks, Gemma 4 31B supports image understanding with variable aspect ratios and resolutions, and can output structured bounding boxes for UI element detection, making it useful for document parsing and UI understanding. Compared to Gemma 3, it delivers stronger reasoning and multimodal performance. It is part of a four-size family alongside the 26B A4B MoE variant and two on-device models (E2B, E4B), with the 31B dense variant optimized for output quality and fine-tuning over inference speed.
GPT-5.6 Luna is the fastest and most cost-efficient model in OpenAI's GPT-5.6 family, which also includes Sol (the flagship tier) and Terra (the balanced mid-tier). Introduced under a new naming convention where the generation number (5.6) and a durable capability tier name (Luna, Terra, Sol) together define each model, Luna occupies the lightweight end of the family and is designed for high-volume, latency-sensitive workloads such as summarization, drafting, autocomplete, classification, and routine automation. The GPT-5.6 family as a whole advances capabilities in software engineering, computer use, professional knowledge work, scientific research, and cybersecurity, with all three tiers rated at the "High" capability level under OpenAI's Preparedness Framework for both cybersecurity and biological/chemical risk domains.
GPT-5.6 Luna supports multimodal input and function calling, and shares the family's 1.5 million token context window. On Terminal-Bench 2.1, Luna scores 82.5%, and on the Artificial Analysis Coding Agent Index it outperforms comparable models at roughly one-quarter the estimated cost of higher-tier alternatives. Luna supports the GPT-5.6 prompt caching scheme, which introduces explicit cache breakpoints and a 30-minute minimum cache life. The model was previewed on June 26, 2026 to a limited group of trusted partners via the OpenAI API and Codex, with general availability rolling out on July 9, 2026 across ChatGPT, Codex, and the API.
Frequently Asked Questions
On Roboflow's Vision Evals, GPT-5.6 Luna performed better. It scores higher on 4 of the six vision tasks and averages 73.8% (#24 of 61) against 67.0% (#38 of 61) for Gemma 4 31B. These scores measure vision capabilities only; either model may perform differently on work outside this benchmark's tasks.
No. On the Vision Evals Counting benchmark at low effort, GPT-5.6 Luna leads with 67.1% against 51.4%. This is the widest gap between the two models across the benchmark's tasks.
GPT-5.6 Luna is cheaper to run. Across Roboflow's Vision Evals task mix it averaged $0.0010 per sample against $0.0015. Actual costs depend on your image sizes, prompts, and output length.
GPT-5.6 Luna is faster. Across Roboflow's Vision Evals it averaged 7.4s per inference against 34.4s. Latency varies with image size, prompt length, and provider load, so treat these as relative rather than guaranteed figures.
Yes. The comparison demo on this page runs both models on the same image side by side for image captioning and OCR in the free Roboflow Playground. You can try it instantly, and a free account unlocks unlimited runs.