{"slug": "gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test", "title": "GPT-6.2 Astra tops 3D benchmark, but Gemini 4 Carbon leads floor plan test", "summary": "OpenAI released GPT-6.2 Astra on November 17, 2026, alongside the lighter GPT-6.2 Sol, and Astra scored 1,318 Elo points on the AA-3D benchmark — the highest mark recorded — while cutting API prices 20% to $8 per million input tokens and $40 per million output tokens. On Blueprint-Bench 2, which tests floor-plan reconstruction from room photos, Google's Gemini 4 Carbon held first place as of November 19, 2026, with a normalized connectivity score of approximately 58.6%. Astra also posted roughly 78% task success on Agents' Last Exam and about 77% on Terminal-Bench-Science 0.1, and ships with a 1.05 million token context window.", "body_md": "# GPT-6.2 Astra tops 3D benchmark, but Gemini 4 Carbon leads floor plan test\n\nPhoto: Tim Witzdam / Pexels\n\nOpenAI's new model leads AA-3D and cuts API prices, while Blueprint-Bench 2 standings show Google's Gemini 4 Carbon in front\n\n[OpenAI](https://cryptobriefing.com/markets/openai/)’s newest flagship model is collecting benchmark trophies. The question is which shelf they belong on.\n\nA post circulating on X credited GPT-6.2 Astra with first place on Blueprint-Bench 2, a test of whether an AI can draw a floor plan by looking at photos of a room’s interior. Standings as of November 19, 2026, however, put [Google](https://cryptobriefing.com/markets/alphabet/)’s Gemini 4 Carbon at the top of that leaderboard, with a normalized connectivity score of approximately 58.6%.\n\nWhere Astra does hold a clear 3D win is the AA-3D benchmark, which tests how well models build interactive 3D experiences. There, Astra scored 1,318 Elo points, the highest mark recorded and ahead of competing models.\n\n## What OpenAI actually shipped\n\nOpenAI released GPT-6.2 Astra on November 17, 2026. A lighter sibling, GPT-6.2 Sol, arrived alongside it, though it offers less capability than Astra.\n\nThe pitch is aimed squarely at professionals. OpenAI is targeting work in science, law and design, and it is leaning hard on efficiency gains as the main selling point.\n\nThe most tangible change for developers is the price. API costs fell by 20% compared with the previous model, landing at $8 per million input tokens and $40 per million output tokens.\n\nAstra also ships with a 1.05 million token context window. That is the amount of material the model can hold in its working memory at once, which matters when you are feeding it an entire contract, a codebase or a stack of research papers.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nAccess is available to ChatGPT Plus, Pro, Enterprise and Edu subscribers, as well as through the OpenAI API.\n\n## The benchmark scorecard\n\nBeyond 3D work, Astra reportedly leads on a pair of agent-focused tests. It posted an approximate 78% task success rate on Agents’ Last Exam and roughly 77% on Terminal-Bench-Science 0.1.\n\nSpatial reasoning is the weaker spot. Astra faces challenges across several spatial-reasoning benchmarks, and Blueprint-Bench 2 is the most visible example.\n\nGemini 4 Carbon’s lead of approximately 58.6% on that test also tells you the ceiling remains low across the industry.\n\nIndependent evaluators paint a similarly mixed picture at the top level. Artificial Analysis places Astra in a competitive position against models like Claude Fable 5.2, particularly on composite intelligence metrics, but not as the outright winner across the board.\n\n## What this means\n\nThe split results on 3D tasks carry a more practical lesson. Astra’s AA-3D lead suggests it is strong at generating interactive 3D content, which could appeal to designers and developers building visual tools. Its weaker showing on floor plan reconstruction suggests that interpreting real physical spaces from photos is a different skill, and one where Google currently holds the edge.\n\nFor the broader AI sector, the takeaway is that no single lab is running away with the field. OpenAI leads on agents and 3D generation, Google leads on at least one spatial reasoning test, and [Anthropic](https://cryptobriefing.com/markets/anthropic/)’s Claude Fable 5.2 remains in the conversation on composite measures.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test", "canonical_source": "https://cryptobriefing.com/gpt-6-2-astra-3d-benchmark-results/", "published_at": "2026-11-20 10:10:21+00:00", "updated_at": "2026-11-20 10:42:09.604591+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-products", "ai-agents", "generative-ai"], "entities": ["OpenAI", "GPT-6.2 Astra", "GPT-6.2 Sol", "Google", "Gemini 4 Carbon", "Anthropic", "Claude Fable 5.2", "Artificial Analysis"], "also_reported_by": [], "takeaway": "GPT-6.2 Astra leads 3D generation at 1,318 Elo on AA-3D but trails Gemini 4 Carbon's ~58.6% on Blueprint-Bench 2 floor-plan reconstruction, so spatial reasoning from photos remains a separate, weaker capability from interactive 3D generation.", "recommendation": {"advice": "Evaluate spatial-reasoning tasks separately from 3D generation: Astra's 1,318 Elo AA-3D lead did not carry over to Blueprint-Bench 2, where Gemini 4 Carbon led at ~58.6%.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test", "markdown": "https://wpnews.pro/news/gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test.md", "text": "https://wpnews.pro/news/gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test.txt", "jsonld": "https://wpnews.pro/news/gpt-6-2-astra-tops-3d-benchmark-but-gemini-4-carbon-leads-floor-plan-test.jsonld"}}