Google's most powerful AI "Carbon" has been unlocked globally, recreating the historic moment of humanity's first moon landing.
Google's most powerful game-changing model to date is fully unlocked!
Demis Hassabis officially announced that the new flagship Gemini 4 Carbon, is now open to users worldwide!
Paid API developers, enterprises and Google AI Ultra users can now open AI Studio, Gemini app and Antigravity to start experiencing it right away.
Two weeks ago, Google already released its debut performance. At that time, it landed straight at the top of the Artificial Analysis rankings with a score of 59.
On multiple leaderboards, Gemini 4 Carbon outperforms Fable 5.2, Opus 5.5, GPT-6 Astra and GPT-6.1 Sol.
In the GraphWalks test, Carbon achieved an ultra-high score of 96.2%. In three real production evaluations including CursorBench 4.0, Perplexity WANDR and Zapier, it took all first places at one go, outperforming Fable 5.2, while its launch API price is as affordable as that of GPT-6.1 Sol!
The team gave Carbon old photos and blueprints of the Apollo Lunar Module, and it fully reconstructed the entire lunar module in 3D space!
Both the complete module body and the dashboard inside the cabin are clearly visible, which is extremely stunning.
3D reconstruction process of the Apollo Lunar Module
Even more stunning is the ace feature launched by Google on the Antigravity platform — /swarm multi-agent swarm mode.
With just one prompt, Carbon can bring the entire AI team to work overnight!
The True Path to ASI
Google's intelligent agent swarm fully embodies the concept of cooperation among tiny intelligent agents.
The most stunning breakthrough this time is that with the newly released AlphaProof Code framework from DeepMind, the AI agent swarm generated 6.5 million lines of proofs in just four weeks, verifying the post-compilation code of Signal's core encryption!
This would originally take human experts several years to complete.
The actual measurement of large-scale construction projects is also very impressive.
This time, the intelligent agent swarm completed the restoration project of Google's Platform 37 office building, including 7 documents and 206 drawings.
Engineers only need to input /swarm and a prompt before going to bed, and a 3D interactive building will be generated!
With only one prompt, the Carbon swarm reproduced the entire construction process of the Eiffel Tower.
It built an open-world game where a pelican rides a bicycle over the roof of the Mountain View headquarters!
It restored the 61-key mechanical structure of the 1780s fortepiano and placed it in the 3D model of Toronto's Koerner Hall, where real-time reverberation can be heard.
It can render underwater Atlantis blockbusters in real time.
Even if you give it a materials science paper, it can easily create a set of fabric mechanics simulation tools to display the stress changes of 110 fibers in real time!
Extreme Real Device Test, Directly Output Black Hole and Minecraft
We also launched the real device test as soon as possible.
It generated the "Gargantua" black hole purely using the general relativity formula.
It can also simulate the ink diffusion on rice paper in real time.
Then we raised the difficulty again, and asked it to create a 3D open world with one sentence.
In the PK between Carbon and Fable 5.2, the demos are similar, but Carbon saves nearly two-thirds of the cost compared to Fable 5.2.
Multi-agent Swarm Works Overnight, Single Task Cost Drops by 6 Times
At present, Carbon costs only $2 per million tokens for input and $10 per million tokens for output, and you can also enjoy a 95% discount for caching.
While taking first place in the three major production evaluations, the average single task cost is more than half lower than that of the second place!
In the multi-modal knowledge agent benchmark MKA released by Google in partnership with Artificial Analysis, Carbon ranked second in the world with a high score of 54.3%.
In the independent anti-ranking manipulation test KillSwitch-Bench recently released by Dominik Picheta, father of the Nim programming language, Carbon and Fable 5.2 tied for first place globally.
In the real financial investment research assessment covering 23 cutting-edge models from hedge fund giant Balyasny, it also took the first place across all participants.
Cloud security unicorn Wiz used Carbon to scan two mainstream SaaS service providers, and found real zero-day vulnerabilities in less than 2 hours.
Security risks that would originally take weeks to check can now be fully completed in one afternoon, protecting tens of thousands of downstream enterprises before hackers take action!
"Free Gift Package" for Global Scientists
This time, Google announced an exciting plan: it is providing free access for up to half a year to 10,000 top scientists around the world!
Whether it is searching for room-temperature superconducting materials, cracking complex protein folding to cure cancer, or simulating plasma turbulence in controlled nuclear fusion, what scientists need is a "research partner that can think and reason".
By providing half a year of free highest-level access, Google hopes these 10,000 scientists in the fields of physics, chemistry, biology, materials science, climatology and other fields can work on hypotheses that have been shelved due to excessive complexity.
In the CASP17 protein prediction blind test, this tool has already helped scholars increase the retrieval volume of difficult viral protein homologs by 40 times.
This time, Google wants to make "giving a prompt before going to bed and checking the results after waking up" a daily routine for scientists!
References:
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
https://x.com/GoogleAIStudio/status/2120187576489222302
https://x.com/GoogleDeepMind/status/2120219371473617099
https://x.com/GoogleAI/status/2120187753329791341
This article is from the WeChat Official Account "AI Era" (ID: AI_era), written by ASI Revelation, and published with authorization from 36Kr.