Just now, Google DeepMind announced that Gemini 4 Carbon is officially available today, billing it as “Google’s most capable model to date.” It is designed specifically for complex, long-horizon tasks, capable of allocating more inference resources to explore problems in depth and complete complex tasks in software engineering, cybersecurity, knowledge work, and scientific research through continuous execution, checking, and correction.
Compared with the earlier Gemini 4 Argon, Carbon has a more clearly defined role:it is not content with quickly providing an answer; instead, it is willing to devote more time and computational resources to truly finish a complex undertaking.
Demis Hassabis gave a tangible example: “Carbon reverse-engineered the binary of Theme Park, the game I originally developed, added a DeepMind-themed attraction to the game, and ported the entire game to the web in one go.”
Below, he also included a video comparison from around the same time last year using Gemini 3, and remarked:“In less than a year, Gemini has evolved from 3 to 4. The progress is truly astonishing…
Seeing all of this happen with my own eyes is truly incredible!”
And now, this powerful capability is available to users worldwide! According to Google, starting today, Gemini 4 Carbon will be available through the following channels:
Consumers: Google AI Ultra subscribers can access it through the Gemini app and AI Mode in Google Search.
Developers and businesses: Developers can use the APIs in Google AI Studio and Android Studio; Google AI Ultra subscribers can also access it through Google Antigravity, while business users can use it through Gemini for Business.
Scientists and researchers: They can apply through Google Antigravity for the opportunity to receive six months of free access.
Cybersecurity defenders: They can apply to join the Fairwind program and gain access to Carbon for cybersecurity use.
In addition, to demonstrate Gemini 4 Carbon’s ability to handle complex scientific research tasks, immediately after the model’s official launch, Google DeepMind presented a representative example and announced a research projectcalled AlphaProof Code: Researchers used an AI agent powered by Gemini 4 Carbon to formally verify the underlying code of the Signal encryption protocol, ultimately generating more than 6.5 million lines of machine-checked mathematical proof code.
This work, which might otherwise have taken a team of experts years to complete, took only four weeks.
More importantly, this means AI is beginning to move from “writing code” toward “using mathematics to prove that code is secure.”
From long-horizon reasoning to actually completing complex work
Google positions Gemini 4 Carbonas the model currently best suited to handling high-difficulty, multi-step tasks.
In software engineering, it can tackle complex codebases, troubleshoot issues across modules, adjust architectures, and conduct multi-round debugging; in scientific research, it can continually explore hypotheses around open questions, analyze evidence, and validate results; in knowledge work, it can also process large volumes of complex documents, data, and structured information.
An important design underlies this:Extended Verification Loops (extended verification loops).
When faced with complex tasks, Carbon devotes more Thinking Tokens to exploring different solution paths and repeatedly checks results throughout execution. When it finds problems, it can continue modifying and testing until it achieves more reliable results.
In the independent workflow benchmarks published by Google, Carbon ranked first in each of CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. Google also said that, in these tests, its per-task cost was less than half that of the next-best-performing external model.
At the same time,Google is previewing the /swarm feature in Google Antigravity: users need only provide a task to have multiple Agents form a collaborative team, separately responsible for architecture design, task coordination, code implementation, review, and other work.
Carbon can take responsibility for the stages that require the deepest reasoning, while other, lighter-weight models handle specific tasks in parallel.
From model capabilities to multi-Agent collaboration, Google hopes to make the work AI performs more complex and closer to real professional workflows.
6.5 million lines of mathematical proof: an AI Agent gives Signal a “security checkup”
If Gemini 4 Carbon showcased the upper limits of capability, then AlphaProof Code is a very concrete validation of that capability.
This research targets a difficult problem that has long troubled the software security field:How can we prove that software actually running truly satisfies the security properties required by its designers?
Take Signal as an example.
As a widely used encrypted communications protocol, Signal’s security is built on a series of complex cryptographic mechanisms. In the past, cryptographers could mathematically analyze whether a protocol’s design was secure, and could also find vulnerabilities through code reviews and testing.
But a gap has always remained: a theoretically secure protocol does not mean that the code ultimately running on users’ devices is necessarily secure.
Between protocol design and actual operation, there are also stages of code implementation, compilation, and optimization. A tiny implementation error can cause an otherwise rigorous security design to fail.
Traditional formal verification (Formal Verification) seeks to solve this problem with mathematical proofs: express the security conditions the software must satisfy as rigorous mathematical propositions, then prove that the actual code indeed satisfies those conditions.
The problem is that this work is extremely complex and often requires experts to devote a great deal of time to it.
AlphaProof Code, which Google DeepMind has now introduced, aims to lower this barrier with the help of AI Agents.
TAG125@@The research team first worked with the model to translate Signal’s security goals into mathematical specifications that Lean could understand, with human experts reviewing each one. Then, multiple Agents powered by Gemini 4 Carbon began automatically constructing proofs, attempting to demonstrate that the actual code satisfied these security requirements.The key here is that the research did not stop at the protocol-design or source-code level; it went further, into the final compiled device code.
The researchers established a formal semantics for the AArch64 instruction set and brought compiled machine code within the scope of verification. This meant that they were trying to prove the correctness of the code actually running on the devices, not merely a theoretical description of the algorithm.
Ultimately, AlphaProof Code generated approximately 6.5 million lines of Lean proof code for Signal’s core implementation, built on approximately 26,000 lines of reviewed formal specifications and machine-code semantics.
The entire verification process took only four weeks. The AI Agent also automatically flagged the portions that required further human review. In the end, only about 1% of the codebase needed additional manual attention, while all the remaining proofs were machine-checked by Lean’s trusted kernel.
More importantly, this system is not merely capable of generating proofs for correct code.
The research team also deliberately reintroduced two vulnerabilities that had genuinely existed in Signal in the past but had since been fixed into an isolated version of the Signal code. As a result, AlphaProof Code successfully generated machine-checked counterexample proofs in both cases, identifying that the code failed to satisfy the original security requirements.
This demonstrates that it can not only help verify correct implementations, but also identify particular incorrect implementations.
Is the next step in AI programming to make code trustworthy?
The significance of this research extends far beyond helping Signal conduct a security verification.
In the past, formal verification has always been one of the most costly security-assurance methods in software engineering. Even for highly critical systems, proving that their code satisfies strict security properties has often required teams of experts to spend years.
Now, AI Agents are beginning to take on much of this arduous work:writing formal specifications, exploring proof paths, generating Lean code, and then having it checked by an independent proof kernel.
Of course, this does not mean that AI can already independently prove that any software is absolutely secure. Human experts are still needed to define and review security specifications, and proof systems can only guarantee that code satisfies properties that have been explicitly articulated.
Nevertheless, this approach has demonstrated a new possibility: when Agents can reason continuously, invoke tools, construct complex intermediate results, and repeatedly check their work through external verification mechanisms, AI has the opportunity to take on work that previously required expert teams over long periods of time.
Google DeepMind has also said that it hopes to expand this technology in the future to critical infrastructure such as medical-device firmware and power-grid software, and to combine it with the automated vulnerability-remediation tool CodeMender, creating a more complete security loop in which software moves from vulnerability discovery and remediation to formal verification.
From the release of Gemini 4 Carbon to AlphaProof Code’s verification of Signal, Google is demonstrating the same direction: the value of models is extending beyond answering questions to undertaking complex work.
And when AI can not only write code, but also help humans use mathematics to prove that the code satisfies security requirements, the boundaries of Agent capabilities are being pushed into a new domain.
What about you—how do you view the fact that today’s large models are no longer limited to leaderboard scores, but are exploring more complex mathematics and cryptographic problems? Feel free to leave a comment and share your thoughts!
Reference links:
https://x.com/GoogleDeepMind/status/2120219371473617099
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=
https://x.com/sundarpichai/status/2120207062655549933
https://deepmind.google/blog/introducing-alphaproof-code/
