Google's Carbon model has entered testing, showing exceptionally strong 3D capabilities and following a recursive self-improvement path. Just eight days after Argon was fully opened to the public, Google's next trump card has already stormed into the arena.
Just moments ago, Google's most powerful model, Carbon, was revealed to be under testing. Partners have already gotten early access, and blind testing has begun on LMArena, with a release seemingly imminent. As soon as the news broke, testers flooded into the Arena. Within a single day, Carbon hands-on results exploded across X, with some users completing an entire level of Super Mario in one go. The excitement over the previous generation flagship has barely cooled, and the next generation is already queuing at the door.
Google co-founder Sergey Brin said that Google has begun stepping onto the "RSI treadmill," where models are helping build the next generation of models.
Super Mario Cleared in One Run, Carbon's 3D Power Is Absurdly Strong
On the Arena, Carbon still carries the name "gemini-4-argon." Testers figured out how to identify it in a rather amusing way 鈥?simply by how slow it is. User Harshith discovered that there are actually two models under this name. One is fast and decent, while the other is surprisingly slow but produces truly stunning work. The latter is Carbon.
Evaluation account Lumina used a Super Mario prompt to have Carbon generate an entire level of the game. The protagonist goes through hitting bricks, collecting coins, and entering pipes, descends into an underground level and climbs back out, finally jumps onto the flagpole and walks into the castle, with "COURSE CLEAR" popping up on screen, along with coin count and rating tallied up.
Lumina also gave the same voxel pagoda task to Argon, Carbon, Opus 5.5, and Sonnet 5.5 for a side-by-side comparison. In Lumina's view, Carbon's quality is clearly higher than Argon's, though depending on the task, it takes 2 to 4 times longer.
User Synopsis's voxel valley pushed the visuals up another level. Cherry blossoms bloom along the riverbank, a red arch bridge spans the water, wooden houses have their lights on, and a rainstorm suddenly begins, darkening the entire valley.
User Salio built a 3D castle where lighting can switch between dusk, blue hour, moonlight, and misty morning light, with the stone bridge below continuously lit.
Well-known leaker chetaslua handed Gemini's mascot to Carbon and asked it to create a 3D animation explaining why carbon is stable. At the start, the mascot stares at a carbon-12 atomic nucleus. Then six electrons arrange themselves as 1s²2s²2p², orbitals hybridize into four identical sp³ lobes, which then share electrons with neighbors to reach eight, finally assembling into a complete diamond lattice. Chetaslua said this task was extremely difficult for previous Gemini models.
Using Carbon to explain carbon 鈥?Google's naming choice this time has been thoroughly appreciated by netizens.
Developer Pankaj Kumar had it create an interactive 3D crystal citrus. However, when it comes to SVG, evaluations began to diverge.
User Binkg asked Carbon to draw an SVG of a cola bottle, with water droplets all over the bottle body, small text on the label and the bottle cap all rendered, with adjustable frost, bubbles, and backlighting on the side. But after watching a full day of testing, Haleemah said Carbon's strongest area is still 3D, and she has yet to see a good SVG.
Beyond SVG, frontend remains the same old problem. Lumina said Carbon's frontend design and aesthetics are still mediocre, lacking that distinctive flavor of Anthropic models. The same model can surge in 3D while standing still in frontend because the two types of tasks are evaluated differently. In 3D scenes, geometry, lighting, and collision are all determined by code 鈥?one calculation error and the visuals collapse, and correctness can be verified at a glance. Frontend tests how much whitespace, what font size, where to place buttons 鈥?there is no standard answer, only human preference. So 3D can be pushed upward through compute and reinforcement learning, while frontend can only be slowly refined through massive human feedback.
Slower but Not Lazier, Carbon Specifically Cures Argon's Laziness
When testers share their work, almost all of them add a note about it being slow, but few complain, because it is no longer lazy. Laziness, ironically, is the most heavily criticized flaw since Argon's launch. Its benchmark scores were beautiful, but within less than a week of developers getting their hands on it, its "lazy" reputation spread. Tech media outlet Every tested Argon and gave five usage recommendations, with the third being to always double-check its work. Out of 38 benchmark runs, 17 times Argon turned the code review step on and then off again, with only 30 milliseconds in between. The same recommendation also contained a set of numbers. Raising Argon from low reasoning mode to high mode only took 1.7 times as long; raising Opus 5.5 to high mode took 18 times as long. In other words, asking Argon to think harder did not make it think much more. On Every CEO Dan Shipper's senior engineer benchmark, both Opus 5.5 and Sonnet 5.5 scored 85%, GPT-6 Astra scored 81%, and Argon scored only 21%, ranking last among 8 models. But with Carbon, the situation flips entirely. Business Insider previously reported that Google employees who tried Carbon internally said writing code with it "feels a lot like Opus 5.5."
Argon Barely Rolled Out Before Carbon Arrived, Google Steps Onto the RSI Treadmill
From Argon to Carbon, Google's pace this round has been astonishingly fast. Argon was officially announced on September 30, initially only given to cybersecurity defenders in the Fairwind program; on October 13, Argon was fully opened to developers and subscribers; just 8 days later, Carbon entered partner testing. Google is playing its cards faster and faster, and Brin's answer is 鈥?RSI (recursive self-improvement). In a recent roundtable discussion, he stated that researchers are now working hand in hand with models to build the next version of the model. Argon's safety report shows that in September of this year, over 90% of DeepMind's code merge requests had more than 75% of the code written by AI. Merge requests fully delegated to AI, with fewer than 3 rounds of human interjection, also rose from 7.5% in April to 11.9%. DeepMind's monthly code merge requests 鈥?the darker the color, the higher the proportion written by AI. When models start helping build models, it is not hard to understand why Argon and Carbon are so close together.
Comments