Qwinah Says Gemini 4 Carbon Is the First Model to Tie Humans on Physical 3D Understanding, Citing a Blueprint Bench 2 Chart
Gemini 4 Carbon is the first model to tie humans in physical 3D understanding we are already approaching AGI, and I can’t imagine what next year holds for us
Context
A model that matches people at reasoning about physical space would help robots and design tools, but this is one poster's reading of a single benchmark chart.
Gemini 4 Carbon is a Google model already covered in the Oct 30 deck through a Perplexity benchmark story. The chart in this post is labeled Blueprint Bench 2 and credited to Andon Labs. The post's wording that the model ties humans is the poster's claim, and the quoted post it links to is about Google AI Ultra usage limits, not the benchmark.
Deep dive
Qwinah (@MaaSonder) wrote on Oct 31 that Gemini 4 Carbon is the first model to tie humans in physical 3D understanding, and added that we are already approaching AGI. The AGI remark is the poster's opinion.
The attached chart is titled Blueprint Bench 2 Scores and carries the Andon Labs name. It shows Gemini 4 Carbon at roughly 0.58, with a dashed line marking human performance at 0.59. Next come Gemini 4 Argon at about 0.54, Claude Opus 5.5 at about 0.51 and GPT 6 Astra at about 0.49, with a long tail of other models below 0.45. By the chart's own numbers, Gemini 4 Carbon sits just under the human line, within its error bar, rather than above it.
The post does not explain what the benchmark measures, how the human baseline was set, or how many tasks it contains, and those details are not confirmed here. The quoted post it links is a separate StatsWire post about Google AI Ultra usage limits in Antigravity.
Gemini 4 Carbon already appears in the Oct 30 deck through a Perplexity benchmark story, so this is a second benchmark claim about the same model rather than a new release. It is ranked low because it is secondhand and rests on one chart.
What to expect next
Watch for Andon Labs to publish the Blueprint Bench 2 method and full results, including how the human score was measured and how many people took part. Independent runs by other groups would show whether the near tie holds. If it does, the next question is whether better scores on spatial tests turn into better results for robotics and design work, where Google has been pitching its Gemini models.