• 2 min read
Claude Fable 5 for Vision: Evaluation and Benchmarks
Claude Fable 5 is a strong reasoning model for visual understand but not a state-of-the-art vision model.
Blog
Erik Kokalj
Developer Experience @ Roboflow
Latest Posts by Erik Kokalj
• 4 min read
Cosmos 3: Evaluation for Vision Use Cases
NVIDIA Cosmos 3 offers zero-shot performance on fixed-camera footage across airports, warehouses, and production lines.
• 6 min read
Gemini 3.5 Flash for Vision: Evaluation and Benchmarks
SUMMARY
Gemini 3.5 Flash, released at Google I/O on May 19, 2026, currently tops the Roboflow Vision Evals leaderboard across 67 real vision prompts covering defect detection, document understanding, object counting, and spatial reasoning. It outperforms Gemini 3.1 Pro on counting and spatial tasks while running roughly
• 5 min read
Serverless GPU Inference Cost Comparison: Roboflow, GCP, AWS, Azure
SUMMARY
Serving a custom RF-DETR XL model on serverless GPU infrastructure produces dramatically different monthly costs depending on the provider and traffic pattern, so this post benchmarks Roboflow Serverless, GCP Cloud Run, AWS SageMaker, and Azure Serverless GPU across three workloads: continuous inference at one request per 10 seconds,
• 5 min read
Which is the Best Coding Agent for Vision tasks?
SUMMARY
This benchmark pits four coding agents (Claude Code with Opus 4.6, Cursor with Composer 2, Gemini CLI with Gemini 3.1 Pro, and Codex with GPT 5.4) against five computer vision tasks including bird counting with SAM 3, car counting in video and RTSP streams, avocado detection,
• 2 min read
Launch: Workflows Builder 2.0
Workflows Builder 2.0 is here with a drag-to-connect interface that makes building and understanding computer vision pipelines faster and clearer.
• 5 min read
Building an AI-powered Bot to Beat the Chrome Dino Game
SUMMARY
This tutorial walks through a complete computer vision project that trains an RF-DETR model to detect the Dino, Cactus, and Bird objects in the Chrome Dino game, then uses model predictions at inference time to drive keyboard actions via pynput. Dataset images were captured by screen-recording manual
• 5 min read
Train a YOLO26 Instance Segmentation Model with Custom Data
SUMMARY
This tutorial walks through fine-tuning YOLO26 for instance segmentation on the Car Part Detection dataset from Roboflow Universe, using a Google Colab notebook with the Ultralytics library. Unlike object detection bounding boxes, instance segmentation identifies the exact pixels belonging to each object, which matters for applications like medical
• 4 min read
How to Train a YOLO26 Object Detection Model with Custom Data
SUMMARY
This tutorial covers fine-tuning a YOLO26 object detection model on the Basketball-Player-Detection dataset from Roboflow Universe, which contains 654 annotated images of NBA players, referees, and basketballs. Using a Google Colab notebook with the Ultralytics library, the guide walks through GPU environment setup, dataset download via
Stay Connected