Google's next flagship Carbon spotted in LMArena blind tests: stunning 3D, fixes Argon's laziness
新智元 · wechat · 2026-10-22
Just 8 days after Argon's full rollout, Google's stronger next model Carbon has reportedly entered testing — partners already have it, and it's blind-testing on LMArena under the name "gemini-4-argon". Testers identify it by how slow it is.
Hands-on results show standout 3D generation: one tester built an entire playable Super Mario level, others made voxel scenes with dynamic weather and switchable lighting, plus a 3D animation explaining carbon stability. Quality is widely judged above Argon despite being 2-4x slower. SVG and frontend aesthetics remain mediocre — likely because 3D outcomes are verifiable and RL-trainable, while design taste needs human feedback.
The contrast with Argon is stark: Argon scored just 21% on Every's staff engineer benchmark (Opus 5.5/Sonnet 5.5: 85%) and was caught toggling code-review steps on and off within 30ms. Google employees reportedly say Carbon's coding "feels a lot like Opus 5.5." Co-founder Sergey Brin says Google is on the RSI treadmill — DeepMind's safety report shows AI wrote 75%+ of code in over 90% of merged PRs in September.
More from Models
- After OpenAI blocks a security request, tekbog argues: this is why open models matter — tekbog · 2026-10-22
- Asked for their most controversial take on humans, 12 AI models gave the same answer 29 of 35 times — Tgoc · 2026-10-22
- OpenAI blocks cybersecurity request, tells authorized researchers to apply for Daybreak access — tekbog · 2026-10-22
- Pichai says internal results show Gemini 5 will be a big jump over Argon — rohanpaul_ai · 2026-10-22
- RAG explainer: real-time retrieval fixes stale data and hallucinations without fine-tuning — goyalshaliniuk · 2026-10-22
- Chat template reversal experiment finds Mistral Small 3 24B best at human mimicry — EntropyMagnets · 2026-10-22