Visual coding hands-on with GPT-6.2 Astra and Sol: first, 100 ways to draw a dog
Arena.ai released a 44-minute hands-on test in which Peter Gostev compared GPT-6.2 Astra, Sol, and Claude Fable 5.2 across 3D worlds, web pages, games, and data visualizations. One test in the video is called 100 ways to draw a dog. Real-conversation scores from Agent Arena are still on the way.
Arena.ai announced that OpenAI's GPT-6.2 Astra and Sol are now live. Agent Arena scores based on real conversations from the global community will be released soon. Before that, Peter Gostev ran a visual coding comparison of them against Anthropic's Claude Fable 5.2.
The video in the source material is about 44 minutes long. The tests cover 3D worlds, HTML animation, websites, games, and data visualization, wrapping up with SVG detail and 100 ways to draw a dog.
Arena.ai was formerly LMArena, a platform where users compare frontier models and give feedback to form public leaderboards. This time it didn't just publish benchmark scores; it first released a hands-on video.
Video chapters are as follows:
- 00:00 Model and pricing background
- 01:08 3D worlds: buildings, water, and detail
- 05:46 Creative range and style consistency
- 11:48 HTML stories and animation
- 16:45 Visual styles such as clay and crochet
- 21:04 Website UI: structure and repeated layouts
- 27:32 Games: visuals, controls, and bugs
- 35:26 Data visualization
- 37:24 Explaining concepts with animation
- 39:28 SVG detail: the peacock test
- 40:57 100 ways to draw a dog
- 42:16 Summary: Astra, Fable, and Sol
Visual coding demos can easily turn into flashy fireworks. A short clip can make people go wow, but once it goes into a project, repeated layouts and bugs expose problems earlier. The video lists website UI and game bugs separately; that arrangement is more informative than the model names. It cares not only about whether a model can generate something, but also about whether it can deliver reliably.
The pricing note is also worth watching. The costs in the video are based on Astra 6 pricing; the actual Astra 6.2 price is 35% lower. That affects usage decisions more than the model codenames.
One commenter said in a reply that 100 ways to draw a dog is his benchmark. The test isn't complicated, but it shows whether a model can keep a recognizable dog and details across a large number of variations.
Peter's prompts and more information are on YouTube. Agent Arena scores are still on the way; once real-conversation data is available, this comparison will be even more interesting.
发布时间: 2026-11-18 03:26