OpenAI Unveils GPT-5.6 Sol with Major Vision Upgrades and Price Cut
모델 출시 | Tue Aug 18 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 2 sources
OpenAI released three new models—Sol, Terra, and Luna—featuring significantly improved vision performance and a 50% price reduction for Sol.
Analysis
[OpenAI] unveiled three-model GPT-5.6 lineup [1]
- Announced new lineup consisting of three models: Sol
- Terra
- and Luna
- Focused launch stream demonstrations on computer use capabilities
- Highlighted UI agents capable of operating desktop applications along with 3D visualizations
[GPT-5.6 Sol] became OpenAI's best vision model to date [1]
- Best performance among all vision models OpenAI has released
- on the Roboflow VLM benchmark
- Scored 46.2 mAP@50 in object detection compared to GPT-5.5's 13.8 mAP@50
- Terra (44.7) and Luna (43.3) also showed major improvements over GPT-5.5
- Handled titles
- paragraphs
- tables
- images
- and signatures well in document layout detection
[GPT-5.6 Sol] cut prices by 50% [2]
- Announced 50% price reduction for Sol via OpenRouter
[GPT-5.6 lineup] disclosed speed and cost trade-offs across models [1]
- Sol averaged approximately 10 seconds per image at about 2.5 cents per image
- Terra ran at about 6 seconds and roughly 1 cent per image
- Luna was the cheapest at just over 5 seconds and under 0.5 cents per image
- Luna offered the best latency-quality balance in the lineup
- with speeds approaching Gemini 3.5 Flash
[GPT-5.6] revealed stability limitations on high-resolution images [1]
- Instability on images larger than 2000x2000 pixels under low reasoning effort
- Higher reasoning effort improves stability but increases token usage
- latency
- and cost
- Recommended prompting with absolute XYXY pixel coordinates for optimal detection results
- Detection performance drops by about 15 mAP when using incorrect coordinate formats