Surprise Attack on Claude, Google's new flagship Gemini has been fully launched, replicating the OpenCode desktop application overnight

According to a report from Zhidx on November 11, in the early hours of today, Google announced that its most powerful AI model Gemini 4 Carbon is officially open to developers, enterprises, scientific researchers and consumers. Gemini 4 Carbon has reached the cutting-edge level in cybersecurity capabilities and long-range Agent workflows in fields including complex knowledge work, software engineering and scientific reasoning.

Google announces the opening of Gemini 4 Carbon (Image source: X)

Two weeks ago (October 28), Gemini 4 Carbon was released and opened to a group of cybersecurity practitioners through the Fairwind program. After this expansion of access, developers and enterprise users can call the API via Google AI Studio and Gemini for business. Google AI Ultra subscribers can experience this model in Google Antigravity, the Gemini app and the AI mode of Google Search.

In independent benchmarks that measure the actual workflow performance of models, Carbon balances cutting-edge accuracy and the cost of practical deployment. It ranks first in the Cursor evaluation list CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, with the cost per task less than half of the highest-ranked external model, outperforming models such as Fable 5.2 and Opus 5.5.

Performance of Gemini 4 Carbon in the CursorBench 4.0 evaluation (Image source: Cursor)

On the Artificial Analysis Intelligence Index, Gemini 4 Carbon (high) ranks second with 60 points, only trailing Claude Fable 5.2 (max), surpassing a large number of models including Claude opus 5.5 (max), GPT-6 Astra (max) and GPT-6.1 Sol (max).

Gemini 4 Carbon ranks second on the Artificial Analysis list (Image source: Artificial Analysis)

In terms of pricing, the trial API price of Gemini 4 Carbon is the same as Argon, which is $2 per million input tokens (about 13.3 RMB) and $10 per million output tokens (about 66.7 RMB), and the discount will last until November 30. After that, the input and output prices will be $4 (about 26.7 RMB) and $20 (about 133.4 RMB) respectively.

Pricing of Gemini 4 Carbon (Image source: Google)

At present, in the Gemini 4 series, Google still recommends using Argon for most software engineering and knowledge work. Carbon is oriented to more difficult tasks such as complex code debugging and open scientific research, which will consume more tokens to carry out more in-depth reasoning and verify results repeatedly.

Sundar Pichai, CEO of Google, posted: "Carbon made a demo of a pelican riding a bike, so you don't have to test it specially." In the animation, the pelican rides a bicycle in the Google campus. After flying over the ramp, the pelican spreads its wings to maintain balance, then lands smoothly on the ground and continues riding.

After the model was opened, some developers posted on the overseas social platform X that Gemini 4 Carbon has been significantly improved compared with Argon, with strong capabilities, and its thinking mode is different from that of Claude and GPT.

Developers' evaluation of Gemini 4 Carbon (Image source: X)

Some other developers said that Gemini 4 Carbon has made great progress compared with previous Gemini models, but its aesthetics is not as good as Fable and Opus, and more detailed prompts or reference cases are needed for guidance.

Developers' evaluation of Gemini 4 Carbon (Image source: X)

01. Ranked second in multimodal knowledge work evaluation, Google office building is reproduced based on 206 engineering drawings

In scenarios such as law, finance and enterprise operation, Carbon can retrieve and analyze a large number of documents, combine the charts, pictures and tables in them to draw conclusions supported by materials, and then generate spreadsheets, presentations or reports according to the specified structure and format.

In order to investigate the ability of Agent to process complex materials such as engineering blueprints, schematic diagrams and data sheets, Google cooperated with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) evaluation. Google said that Carbon ranked second in this test, and the average cost per task was less than one sixth of the top-ranked model.

Google demonstrated how Carbon turns different types of materials into interactive applications. Among them, a group of Carbon Agents turned historical photos and engineering drawings into a Three.js-based 3D exhibition according to one prompt, showing the construction process of the Eiffel Tower from 1887 to 1889.

In another demonstration, the Carbon team extracted 206 engineering views from a prompt and 7 public planning PDFs released by the city council, and after 10 rounds of iteration overnight, created an interactive 3D model of Google's Platform 37 office building, containing about 906,000 triangular faces.

02. Outstanding multi-Agent collaboration capability, rewriting the OpenCode desktop application overnight

Google said that Carbon is good at engineering tasks that require multi-Agent collaboration, including cross-package migration, cross-subsystem troubleshooting of root cause of faults and complex architecture refactoring. When developing web applications, Carbon will repeatedly build, run and verify, try to recover after encountering errors, and check the source code and the actual rendering effect of the page.

The OpenCode team once gave Carbon a list of requirements before off work, hoping to make the desktop application visually consistent with the existing interface framework. The next day, Carbon had built a Rust interface library from scratch, rewritten the OpenCode desktop application, and delivered a runnable demo. HubX tested Carbon with 16 real bug fix tasks in a codebase containing 14,000 tests, and the result showed a 94% fix rate with no regression during the test.

In the case demonstrated by Google, Carbon took the video generated by Gemini Omni as a reference, built procedural ruins, underwater caustic lighting and water body simulation in Blender, then transplanted the underwater Atlantis scene to Three.js, rendered it in real time in the browser, and generated an interactive 3D game.

03. Leading in vulnerability discovery, exploitation and repair evaluation, with 100% recall rate in Google internal scanning test

Gemini 4 Carbon has the strongest cybersecurity capability Google has ever had, and can provide end-to-end support for security teams to protect critical software. Carbon used through the Fairwind program leads in three key evaluations: vulnerability discovery (SEC-bench Pro), vulnerability exploitation (ExploitGym) and vulnerability repair (CWE-bench v1).

Google tested Carbon in its internal automated security project Project Fortify, which covers scanning, triage and repair. According to Google's internal evaluation, Carbon's scanning recall rate reaches 100%, and the triage precision rate reaches 95%. In the open source codebase test, its vulnerability discovery recall rate is more than 25% higher than that of Argon, and it identifies logical vulnerabilities that were previously missed.

In the "Scan for Good" project of Wiz, a cybersecurity company under Google, the team used Carbon to evaluate the security of public digital infrastructure. Google said that when Wiz inspected two large SaaS service providers, it discovered previously unknown vulnerabilities in less than two hours, and then cooperated with the affected organizations to fix them, with potential impact involving tens of thousands of companies.

04. Participate in verifying the code of the Signal protocol, and the free scientific research program is expanded to Carbon

Google said that Carbon has improved in physical, mathematical and life science reasoning, and can be used to integrate experimental data, test hypotheses and handle multi-step problems. Axiom Bio, a drug toxicity research company, said that Carbon performed best in its translational science tasks, helping the team find previously undiscovered toxicity mechanisms. The company is migrating the analysis workflows and Agents used for actual scientific research to Carbon and Argon.

Inside Google, researchers have used Carbon for formal code verification, enzyme detection method development and protein structure prediction.

In terms of code verification, researchers built an Agent framework called AlphaProof Code for mathematical verification of the actual code of the Signal encryption protocol. With the help of Carbon and other models, researchers transcribed the security goals into mathematical theorems, and Carbon generated more than 6 million lines of Lean automatically checked proof code, shortening the verification work that originally took experts several years to complete to several weeks, and only about 1% of the codebase requires manual review.

In terms of enzyme detection method development, Carbon found unexplainable anomalies based on experimental records and raw data, helped researchers identify problems such as insufficient automatic mixing of deep well plates and residual trace samples in the instrument, and improved the chromatographic peak resolution method. Google said these improvements increased the reproducibility of experimental results and accelerated the development of new detection methods.

In terms of protein structure prediction, models such as AlphaFold 3 can use evolutionarily related protein sequences to infer the folded structure with the help of multiple sequence alignment. For proteins with few known homologous sequences in the database, there is less information available for prediction models. Google said that among the 7 low-homology protein families in the CASP17 blind test set, the number of homologous sequences retrieved with the assistance of Carbon increased to an average of 40 times, increasing the available information for the structure prediction model without retraining it. A broader assessment is still ongoing.

About four weeks ago, Google launched a free program for 10,000 scientists, providing 6 months of access to Google Antigravity, and thousands of scientists have joined. After this opening, both existing and new participants will have the opportunity to use Carbon and Argon. Google will open access in batches every week according to qualification, capacity and screening results, and plans to continue providing participants with access after the new generation of cutting-edge models go online.

05. Conclusion: Carbon is officially opened, and complex tasks enter actual combat test

From rewriting the desktop application overnight to participating in verifying the machine code of the Signal protocol, Google demonstrates how Gemini 4 Carbon repeatedly calls tools, checks results and corrects solutions in one task.

After Carbon goes online, Google has drawn a clearer division of labor for the Gemini 4 series: Argon handles most daily development and knowledge work, while Carbon uses more inference tokens to solve difficult problems that require multiple rounds of verification. With the official opening of Carbon, its task time consumption, token consumption and how much manual review is required before delivery will become the considerations for users to choose models.

This article is from the WeChat official account "Zhidx" (ID: zhidxc), written by Yang Jingli, edited by Li Shuiqing, authorized for release by 36Kr.