Opus 5.5 has won the timeline, not yet the ledger

Why it matters

【Observation】

In the week after Claude Opus 5.5 launched (2026-09-22, https://www.anthropic.com/claude-opus-5-5), X discussion of “Opus 5.5 vs GPT-6 Astra” tilted clearly toward Opus. But most of the volume came from highly shareable visual demos—one-prompt motion-graphics videos (e.g., @stephanlivera https://x.com/i/status/2103315922098470926; @shneural’s same-prompt comparison https://x.com/i/status/2103151003272962130), playable three.js scenes (@MengTo https://x.com/i/status/2102760783344189761), and code-generated videos (@NFT_Chen https://x.com/i/status/2102681172367323300)—rather than controlled, like-for-like comparisons. Reposts and endorsements from Elon Musk (https://x.com/i/status/2104360927474921529) and DHH (https://x.com/i/status/2104273594217812283) amplified the mood. The counter-evidence comes from @rexan_wong’s breakdown (https://x.com/i/status/2103707054108299437): the supposed “one-prompt” videos actually relied on reference videos, rendering frameworks such as HyperFrames/Remotion, ready-made component libraries, multiple storyboard variants, per-scene still frames, and director-style revision notes; @dotey likewise said his piece was “not one prompt, but polished many times” (https://x.com/i/status/2102940980379017293). Chinese-language skeptics reached similar conclusions: @xiaohu warned that “a strong model doesn’t make you strong” (https://x.com/i/status/2104406001948754164); @dingyi argued that without good prompts and source material you get “the typical Claude AI-slop animation” (https://x.com/i/status/2103880309247017445); @fankaishuoai said Astra’s perceived weakness comes from Codex’s rules and safety constraints, and that it performs differently in another agent harness—the model and the harness are two different things (https://x.com/i/status/2101935525489950871); @gengdaJ finds Opus 5.5 stronger but still hands computer-use execution to Codex (https://x.com/i/status/2104405179827359892). Chinese-circle-specific factors also matter: