Claude Opus 5.5 vs GPT-6 Astra — Visual Forge Benchmark

Claude Opus 5.5 and GPT-6 Astra get the same briefs in my VisualForge benchmark: a wind tunnel, a jetpack game, a battle game, a Starship-style rocket flight, a camera shop website, a jet engine, a helicopter game and a bug-fixing app. Every page is scored in a real browser, and every result is public, so you can open the pages and judge them yourself. Compare the two runs: https://visualforgebench.com/compare/?a=20261003T011656_claude-opus-5-5_eca7&b=20261003T110716_gpt-6-astra_10d6 All results: https://visualforgebench.com Average score: Claude Opus 5.5 7.3 out of 10 over eight tests, GPT-6 Astra 7.1 over nine (including the wildcard test it designed itself). Worth knowing before you compare: - The judge is Claude Opus 5.5, so Opus is marking its own homework. Opus's original Alex Optics page was scored by Claude Sonnet 5.5. - Alex Optics: the 5.6 comes from Opus's original code. The repaired page (7.8) is shown for comparison only. - Astra ran on the newer briefs, which ask for sound. Opus ran on the older briefs, without sound. - The bug-fixing test used different apps: Opus fixed Hangar Ops (370 lines), Astra the bigger Hangar Planner (about 900 lines). - Both models ran at high effort on my subscriptions. Costs are what the same work would cost at API prices; Astra's figure is an estimate. I recorded my commentary while going through the results. The narration is a tightened script of what I said, read by my own HeyGen voice clone. 0:00 Intro 0:51 The numbers 1:50 How the scores work 2:37 Aero Lab 4:17 Jet Odyssey 5:33 Blast Island 6:32 Launch Day 9:13 Alex Optics 10:57 Afterburner 12:05 Bug fixing 12:39 Rotor Rampage 13:22 Wildcard: Unzip the Desert 13:43 Close Which model should I test next? Tell me in the comments.