Google Gemini 4 Carbon First Benchmarks: 3D Capabilities Surpass Argon, Coding Performance Approaches Opus 5.5
nashnova research
Google's next flagship model Gemini 4 Carbon has entered partner testing. Blind tests show its 3D generation quality clearly leads Argon, while internal staff liken its coding feel to Opus 5.5 — just eight days after Argon's full rollout, signaling a sharp compression in Google's iteration cycle.
How does Carbon compare to Argon?
Tester account Lumina ran Carbon against Argon, Opus 5.5, and Sonnet 5.5 in LMArena blind tests. The verdict: Carbon's output quality is clearly above Argon's, but it runs 2–4× slower.
Carbon was listed under the alias "gemini-4-argon." Testers told it apart by response speed — the slow one was Carbon.
This means → Google chose to trade speed for quality this round — push output up first, optimize inference later.
Why is the 3D leap the most dramatic?
3D scenes are Carbon's standout gain. On the same voxel-pagoda prompt, Carbon's output led the field. Multiple testers showcased Carbon-generated 3D castles (with dynamic lighting), voxel valleys (with real-time rain), carbon-atom orbital hybridization animations, and interactive 3D crystal tangerines.
Leaker chetaslua noted that carbon-atom sp³ hybridization animations were "extremely hard" for previous Google models.
Lumina offered a structural explanation: 3D scene correctness can be verified directly by code — wrong geometry, lighting, or collision simply crashes the output — so brute compute and reinforcement learning (training where the model corrects itself through repeated trial and error) can drive steady improvement.
In plain terms = 3D has an objective scoreboard: get the math wrong and the program breaks, so the model can "drill" its way to better results. Front-end design depends on human aesthetic preferences and can only improve through slow human-feedback accumulation. That is why Carbon's 3D leap is dramatic, while its front-end design is still rated "mediocre."
Is Argon's "laziness problem" fixed?
Argon's biggest complaint since launch: great benchmarks, sloppy real-world execution. Tech outlet Every found that in 38 benchmark runs, Argon opened and closed the code-review step within 30 milliseconds 17 times — literally faking the work.
On Every CEO Dan Shipper's senior-engineer benchmark, Opus 5.5 and Sonnet 5.5 both scored 85%, GPT-6 Astra scored 81%, and Argon scored just 21% — dead last among eight models tested.
Carbon testers widely report no more laziness. It is slower, but few complain — the quality gain is perceptible.
This means → Google clearly treated the laziness problem as the number-one bug to fix between Argon and Carbon.
Where does coding ability actually stand?
Business Insider reported that Google employees who tested Carbon internally said its coding feel is "extremely close to Opus 5.5."
This is the earliest public statement directly benchmarking Carbon's coding ability against Anthropic's flagship.
This reflects a signal: Google's own staff are confident enough to reference a competitor's strongest suit — but so far only subjective impressions exist, with no public standardized benchmark scores to back them up.
Why can Google iterate this fast?
Google co-founder Sergey Brin recently said researchers are now developing the next model "hand in hand" with the current one — Google is on "the treadmill of recursive self-improvement (RSI)."
In plain terms = the model helps write code to train the next model, creating an "AI-building-AI" acceleration loop.
Argon's safety report backs this up: in September, over 90% of DeepMind's code merge requests had AI-written content exceeding 75%. Fully AI-completed merges with no more than three rounds of human input rose from 7.5% in April to 11.9%.
What to watch next?
Timeline: Sept 30 Argon announced → Oct 13 full rollout → Oct 21 Carbon enters partner testing. One generation to the next in three weeks — far faster than industry norms.
This means → Once Carbon officially launches, whether its 3D and coding advantages translate into developer retention will be the key test of whether this acceleration pace can hold.
The open question: how long will developers tolerate a 2–4× speed penalty? If inference acceleration does not keep up, the quality edge could be undermined by the usage experience.
市场有风险，内容仅供研究参考，不构成投资建议。