Claire Vo's 'Barbie Bench' Exposes AI Spatial Blindspots
Claire Vo explains why Claude Opus 5.5 and GPT-6 fail the Barbie Bench 3D test, proving frontier AI still cannot render basic human anatomy.
Opus 5.5 is great (but GPT-6 Sol is better)
With Claire Vo · Sunday, September 27, 2026
Claire Vo runs a live, blind benchmarking session comparing OpenAI's GPT-6 Sol and GPT-6 Astra against Anthropic's Claude Opus 5.5 across real-world workflows including PRDs, email triage, frontend coding, SVGs, and video editing. Vo evaluates the models on speed, ergonomics, personality, and design aesthetic, contrasting her human ratings with automated LLM-as-a-judge scores. The episode highlights why Claude Opus 5.5 excels at complex B2B frontend and agentic tasks while OpenAI's models dominate in writing clarity, token efficiency, and creative illustration.
Key takeaways
- Standard AI benchmarks track spaceship games and Python scripts, but Claire Vo uses the "Barbie Bench" (generating a functional 3D Barbie fashion designer game) to test spatial reasoning. Read more →
- Public leaderboards measure synthetic trivia, while production work depends on ergonomics, speed, and day-to-day readability. Read more →
- Anthropic fixed the chatty verbosity of Opus 5, but Opus 5.5 stops narrating its actions, which makes execution feel artificially slow. Read more →
- OpenAI's GPT-6 Astra and Sol outperformed Anthropic's Claude Opus 5.5 on creative vector illustration, generating better character SVGs with defined shadows. Read more →
- Claude Opus 5.5 builds data-dense B2B interfaces that excel at complex workflows, but it tends to overcrowd screens on early prototypes. Read more →
- Automated LLM evaluators frequently penalize readable, concise outputs while rewarding structural verbosity and syntactic compliance. Read more →
6 articles from this episode
Claire Vo's Blind Benchmarking Method for LLMs
Claire Vo shares her blind taste test framework for testing frontier AI models like Claude Opus 5.5 and GPT-6 Astra across real workflows.
Why Claude Opus 5.5 Still Loses to GPT-6 Sol on Speed
Claire Vo benchmarks Claude Opus 5.5 against GPT-6 Sol across speed, price, and verbosity in daily product workflows.
Why OpenAI Beats Claude on Vector Art and Fails at Video
Claire Vo blind-tests GPT-6 Astra, Sol, and Claude Opus 5.5 on SVGs and video editing. Here is why the results surprised everyone.
Benchmarking Frontend Code: Claude Opus 5.5 vs. GPT-6
Claire Vo benchmarks Claude Opus 5.5 against GPT-6 Sol and Astra on frontend code, exposing key design biases and UI bottlenecks.
Stop Trusting LLM Judges to Pick Your Foundation Models
Claire Vo found automated LLM judges reward verbose outputs over human readability. Here is how human blind tests expose model evaluation flaws.
More How I AI episodes
- Claude is BACK with Opus 5.5Sep 27 · 4 articles
- Stripe built a company brain: Meet KaiSep 13 · 5 articles
- GPT-6 Astra blew away every one of my benchmarksSep 6 · 5 articles
- 7 Grok Bot agents I use every daySep 6 · 5 articles