TL;DR
The short version
On release day Vo runs a live blind How I AI bench across Opus 5.5, GPT-6 Sol, and GPT-6 Luna. Her gut crowns Astra/Sol for joy and SVGs; Opus 5.5 earns “back in the mix.”
The punchline on How I AI: her GPT-based LLM-as-judge prefers Fable — proof that vibe ≠ automated score.
Distilled from Claire Vo’s live How I AI stream I reviewed Opus 5.5 and GPT-6 Sol live.
Daily drivers, not frontier theatre
Opus 5.5, GPT-6 Sol, and GPT-6 Luna land as cheaper/faster workhorses under Astra/Fable. Vo’s chart had Opus about twice Sol — enough to split “favorite” from “default” once prices posted.
I found myself annoyed zero times with Opus 5.5. And when I thought I was annoyed with Opus 5.5, I was actually accidentally on Opus 5.
Claire Vo
Wider bench, messier truth
Beyond PRDs/prototypes: agent personality, ~86-turn research memos, computer use, SVG characters, video shorts. Consumer UI still broken across labs; B2B/SaaS often cleaner on Claude; OpenAI models surprised on character SVGs; nobody YOLO-pushed to prod.
How to route
Sol/Codex for cheap daily feel and PRDs. Opus 5.5 for long agentic runs, assistant voice, selected front-end — accept the safety personality. Astra for quality spikes. Multi-model adversarial review as habit.
Ignore anyone selling a single winner from a live n=1 stream. Steal the stack design: price tier, personality tax, and judge bias are first-class routing inputs.
