Teaching AI to Nitpick: OpenAI's GPT-6 Astra Reaches 80% Accuracy in Spotting Furniture Assembly Errors from Manuals and Photos
Summary
Epoch AI released a Furniture Assembly Benchmark (FAB) on September 23, using 60 photos of IKEA furniture assembly to evaluate multimodal AI's visual and spatial reasoning. OpenAI's newest GPT-6 Astra topped the ranking at 80% accuracy, followed by Anthropic's Claude Fable 5.1 at 70% and Claude Opus 5 at 61%, while the leading model of November 2025, Claude Opus 4.5, scored only 28%. China's best open-weight model, Kimi K3, trails the international frontier by about seven months, and popular models such as DeepSeek V4 Pro and GLM 5.3 still lack image support.
Bubble analysis
This is a model-capability benchmark story with no direct bearing on whether AI investment is overheating: it discloses no capex, funding totals or valuations, and says nothing about AI revenue versus spending. At most it weakly suggests that frontier model progress is still fast, offering mild evidence that the technology is genuinely advancing rather than being pure hype — but that is nowhere near enough to support or refute a bubble call.
Related signals
linked by shared tags
- OpenAI Says It Was Hit by Organized Distillation, Points Finger at Moonshot AI#kimi#openai
- Kimi K3 Joins OpenAI's Enterprise Paid Ecosystem, Driving GPU Compute Procurement Demand#kimi#openai
- OpenAI's New Essay Pulls Superintelligence Back to the Execution Layer: Even a Billion Einsteins Need Someone to Mine#openai
- Racing to List Before Thanksgiving: Anthropic Said to Seek IPO as Early as Mid-November#openai
- OpenAI's new agent is a shot at Meta — but can it compete with free?#openai