Artificial Analysis Launches AA-3D Benchmark for 3D Scene Generation

A developer on V2EX shared a 7-day public experiment leveraging the newly released Gemini-4-Argon as a startup executive and DeepSeek models as employees. The setup assigns Gemini 4 to handle long-range planning and core engineering, while DeepSeek handles testing and operations within a $1,000 budget.

Artificial Analysis has added DeepSeek V4.1 Flash to AA-AgentPerf, its agentic inference performance benchmark, featuring results across NVIDIA and AMD hardware. At 100 tokens/s per user, NVIDIA GB300 leads with top-tier throughput.

A developer shared an experiment running a software startup using Gemini 4 as management and DeepSeek as staff, following the official release of Gemini 4 and its strong long-horizon planning capabilities.

A Jinjue author argues that complaints about DeepSeek 4.1 Flash being 'dumbed down' come down to how it is used, not the model itself. The post ships production-ready code and cost figures for model routing and async batch tagging, plus a dual-model setup that uses a lightweight model as a verifier.

Artificial Analysis has announced the launch of AA-3D, accompanied by a public leaderboard and comprehensive evaluation results. The benchmark provides an independent comparison framework for 3D generation and related AI model capabilities.

Artificial Analysis has released an interactive AA-3D voting tool that lets users compare two unnamed models generating 3D scenes from the same brief. Participants explore the scenes for 30 seconds before voting to reveal the underlying models.

Four models form the AA-3D cost frontier, each with a higher Elo than every cheaper model. GPT-6 Luna costs $0.20 a scene at Elo 1000, while Claude Sonnet 5.5 costs $15.18 a scene at Elo 1194; the remaining tied top-four models run $18.11 to $32.08 a scene.

A top-five model has the highest win rate in six of AA-3D's seven domains. In the living bodies domain, sixth-ranked GPT-6.1 Sol (78.6%) ties with Claude Sonnet 5.5 (78.5%), and gaps under roughly 12 points are treated as noise.

Artificial Analysis is launching AA-3D, a new benchmark for how well AI models build interactive 3D scenes. Each model works as an agent in an offline sandbox, turning a written brief into a three.js web page that a human panel then judges.

Speaking at a DBS forum, Nvidia CEO Jensen Huang said Chinese AI models including DeepSeek, Kimi and GLM are 'already good enough.' He described himself as a firm supporter of open source, calling it the best way to drive efficiency gains, and added that the market will decide which models fit which use cases.