GPT-6.1 Sol vs Sonnet 5.5 — Visual Forge Benchmark

OpenAI has announced GPT-6.1 Sol, claiming near-Astra intelligence at a fifth of Astra's standard token price. I wanted to see what that looks like in the sort of applications I actually test. I open the saved results from my VisualForge tests: a wind tunnel, a canyon flight game, a battle game, a rocket simulation, a camera shop, and a separate creativity task. I much prefer Sonnet's visual finish in these examples, but Sol does better on the Earth-view requirement in the rocket test. These are my visual and first-use impressions of the builds shown. Run settings and output history are not fully matched. The displayed scores cover different task sets, and Sol's judge reported missing screenshots, so those totals aren't a fair head-to-head benchmark. I haven't tested Astra alongside these builds. Screen recording: 29 September 2026. The narration uses my own HeyGen voice clone to replace the echo from two microphones in the original recording. 00:00 The result 00:37 How I tested 01:25 Aero Lab 02:19 Jet Canyon 03:05 Blast Island 04:02 Rocket simulation 04:48 Alex Optics 05:46 The creativity task 06:34 My verdict OpenAI announcement: https://openai.com/index/introducing-gpt-6-1-sol/ Claude Sonnet 5.5: https://www.anthropic.com/claude-sonnet-5-5