Roboflow

Gemma 4 31B vs GPT-6 Luna

Compare Gemma 4 31B and GPT-6 Luna side-by-side.

Compare Gemma 4 31B vs GPT-6 Luna live

Run the same image across every model that supports a task and compare their outputs side-by-side.

These models don't share enough common tasks for a side-by-side demo. See the comparison table below for their capabilities.

Models in this comparison

Gemma 4 31B vs GPT-6 Luna on Vision Evals

GPT-6 Luna scores higher on 4 of the six Vision Evals tasks.

The widest gap is Counting, where GPT-6 Luna leads 65.8% to 51.4%.

Overall, Gemma 4 31B averages 67.0% (#34 of 57) against 68.6% (#32 of 57) for GPT-6 Luna.

GPT-6 Luna is both cheaper ($0.0004 vs $0.0012 per sample) and faster (11.3s vs 28.8s per sample).

Gemma 4 31BGPT-6 Luna

Gemma 4 31B vs GPT-6 Luna Comparison Table

Evals updated September 22, 2026Pricing updated September 22, 2026

PropertyGemma 4 31BGPT-6 Luna
OrganizationGoogleOpenAI
Categoryopenclosed
Modalitymultimodalmultimodal
Release DateApr 2026Sep 2026
Context Window256K1.1M
Parameters31B
LicenseApache 2.0Proprietary
Pricing per 1M tokens
Input $/1M$0.090
Output $/1M$0.340
Vision Tasks
CaptioningDemo
Chart Question Answering
ClassificationDemo
Document Question Answering
Image Tagging
Multi-Label Classification
Object DetectionDemo
OCRDemo
Vision Language
Visual Question AnsweringDemo
Model Features
Foundation Vision
LLMs with Vision Capabilities
Multimodal Vision
Vision Evalsground-truth scores across 6 vision tasks, pooled at low effort
Overall
67.0%
68.6%
Quantizationsself-hosted
BF1665.3%FP865.1%QAT-W4A1667.0%hardware →
Avg cost / sample$0.0012$0.0004
Avg speed / sample28.79s11.27s
By task
Object Detection (low)
48.2%
±0.2, Mean of 3 runs, range 48.0 to 48.4
$0
56.8%
±1.9, Mean of 3 runs, range 54.8 to 58.5
$0.0006
Object Detection (high)
64.1%
±0.5, Mean of 3 runs, range 63.6 to 64.5
$0.0016
Counting (low)
51.4%
±1.4, Mean of 3 runs, range 50.0 to 52.7
$0
65.8%
±1.4, Mean of 3 runs, range 64.9 to 67.6
$0.0003
Counting (high)
64.4%
±2.0, Mean of 3 runs, range 62.2 to 66.2
$0.0006
Identification (low)
80.2%
±3.1, Mean of 3 runs, range 78.1 to 84.4
$0
81.3%
±0.0, Mean of 3 runs, range 81.3 to 81.3
$0.0002
Identification (high)
80.2%
±1.6, Mean of 3 runs, range 78.1 to 81.3
$0.0003
OCR (low)
90.8%
±0.2, Mean of 3 runs, range 90.6 to 90.9
$0
87.9%
±0.6, Mean of 3 runs, range 87.2 to 88.3
$0.0005
OCR (high)
88.5%
±0.6, Mean of 3 runs, range 87.9 to 89.2
$0.0014
Data Extraction (low)
80.4%
±2.6, Mean of 3 runs, range 77.3 to 82.5
$0
68.0%
±3.1, Mean of 3 runs, range 65.0 to 71.1
$0.0002
Data Extraction (high)
66.7%
±0.5, Mean of 3 runs, range 66.0 to 67.0
$0.0004
Reasoning (low)
50.8%
±1.7, Mean of 3 runs, range 49.0 to 52.3
$0
52.1%
±2.0, Mean of 3 runs, range 49.7 to 53.6
$0.0003
Reasoning (high)
60.7%
±1.7, Mean of 3 runs, range 58.9 to 62.3
$0.0006

Gemma 4 31B vs GPT-6 Luna: Overview

Gemma 4 31B

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-6 Luna

GPT-6 Luna is the fast, cost-efficient tier of OpenAI's GPT-6 model family, sitting below GPT-6 Sol and the larger GPT-6 Astra model that opened the generation. It is a proprietary multimodal transformer that accepts text and image input and returns text, and it exposes an adjustable reasoning effort setting so the same model can run in a low-latency mode or spend additional inference compute on harder problems. OpenAI positions it for high-volume and latency-sensitive workloads such as conversational assistants, classification, and lightweight agentic pipelines, while noting that at higher reasoning effort it handles software engineering and computer-use tasks that previously required a Sol-tier model.

The model supports a context window of roughly 1,050,000 input tokens with a maximum output of 128,000 tokens, which allows long documents, extended agent traces, and large code repositories to be processed in a single request. OpenAI describes the GPT-6 generation as improving factual reliability and adopting a more concise communication style relative to the GPT-5.6 series, and attributes the efficiency of the Sol and Luna tiers to gains in caching and inference rather than to reduced capability. Architecture details, parameter counts, and training data are not published.

Frequently Asked Questions

On Roboflow's Vision Evals, GPT-6 Luna performed better. It scores higher on 4 of the six vision tasks and averages 68.6% (#32 of 57) against 67.0% (#34 of 57) 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-6 Luna leads with 65.8% against 51.4%. This is the widest gap between the two models across the benchmark's tasks.

GPT-6 Luna is cheaper to run. Across Roboflow's Vision Evals task mix it averaged $0.0004 per sample against $0.0012. Actual costs depend on your image sizes, prompts, and output length.

GPT-6 Luna is faster. Across Roboflow's Vision Evals it averaged 11.3s per inference against 28.8s. Latency varies with image size, prompt length, and provider load, so treat these as relative rather than guaranteed figures.