SI Agent News
DevelopingGeminiAgentsSecurityResearch

Google opens Gemini 4 Carbon to scientists and cyber defenders

Google is providing free access to its new high-reasoning AI model for researchers to tackle complex problems in biology, math, and cybersecurity.

40xincrease in related protein sequences retrieved

The short version

Why it matters

This provides specialized AI tools to researchers who need more than just a chatbot; they need a system that can handle complex logic and long-form verification. It could significantly speed up the pace of drug discovery and the identification of software vulnerabilities.

Expanding access for research

Google DeepMind is expanding its scientist access program to include its newest model, Gemini 4 Carbon. This initiative is part of a broader effort to provide 10,000 scientists with six months of free access to advanced research tools. Thousands of researchers have already joined the program, which originally only offered the Gemini 4 Argon model.

Current and future participants will now be able to use Carbon for their work. Applications for this access remain open, and Google selects researchers on a weekly rolling basis. The program is hosted through Google Antigravity, a platform designed to provide high-usage tiers for specialized research.

While the research program offers free access, Carbon is also being released to paying developers and enterprises on November 10. This includes Google AI Ultra subscribers and those using the Gemini API. The move signals a shift toward making high-reasoning models available for both academic exploration and commercial production.

Deep reasoning and verification

Gemini 4 Carbon is positioned as a tool for demanding tasks that require extended reasoning and verification. Unlike the Argon model, which Google recommends for standard software engineering and knowledge work, Carbon is built for more complex problems. It is designed to explore thoroughly and run extended verification loops to ensure accuracy in its outputs.

The model is particularly useful for open-ended scientific investigation and complex coding. It can handle multi-step processes where the AI must check its own work or verify data against specific constraints. This makes it a suitable candidate for debugging difficult software issues or investigating raw experimental data from laboratories.

Google reports that Carbon shows high levels of honesty in its internal benchmarks. In a test of 375 tasks involving agent honesty, the model passed 92% of the time. It also demonstrated a high rate of identifying methodological flaws and frequently declined to fill in missing experimental data to maintain integrity.

Applications in biology and math

Researchers are already using Carbon to tackle difficult problems in protein prediction and laboratory data analysis. For example, the model was used to search archives for distant evolutionary relatives of proteins that are hard to model because they lack sequences in standard databases. In these tests, the model increased the number of retrieved sequences by an average of 40 times.

These additional sequences provide more information for structure prediction systems like AlphaFold 3. Beyond biology, the model has been used to identify anomalies in laboratory notebooks, such as incomplete automated mixing. It has also helped improve methods for separating chromatography peaks and detecting sample carryover in research instruments.

In the pharmaceutical space, the company Axiom Bio is moving its production scientific analyses to Carbon and Argon. Their testing with these models identified toxicity mechanisms that the company had not previously discovered. This demonstrates how the models can assist in identifying nuanced patterns in complex biological data.

Advancing cybersecurity and code

The model is also being used to accelerate software verification and cybersecurity defense. Researchers combined Carbon with other models in a framework called AlphaProof Code to translate security objectives into mathematical theorems. This process generated millions of lines of Lean code to establish security properties for Signal’s cryptographic protocol.

Google reports that this work was completed in weeks, whereas similar formal verification tasks have historically taken expert teams years to finish. Additionally, the cybersecurity firm Wiz used Carbon to identify previously unknown vulnerabilities across two major software-as-a-service providers in under two hours. These findings allowed the affected organizations to begin remediation efforts quickly.

Google’s Cloud Vulnerability Research team also used the model to find exploitable bugs in open-source software. The maintainers of that software subsequently issued patches for the critical vulnerabilities. These applications highlight Carbon's ability to evaluate complex codebases and determine if a bug is actually exploitable by an attacker.

New tools and pricing models

Alongside the new model, Google is previewing a /swarm feature within the Antigravity platform. This feature allows one prompt to trigger a coordinated group of agents, where different agents can be assigned specific roles like architectural planning or auditing. This allows developers to use Carbon for heavy reasoning while using faster, cheaper models for simpler tasks.

The system also features improvements in handling dense documents, including charts, images, and tables. It can generate various outputs such as spreadsheets, presentations, and reports. In a benchmark by Artificial Analysis, Carbon ranked second for its performance while costing less than one-sixth of the average cost of the highest-performing model.

For paying users, Carbon shares the same introductory API pricing as Argon, which is $2 per million input tokens and $10 per million output tokens. These options are available through the Gemini API in Google AI Studio and Android Studio, as well as through enterprise routes like Gemini for business and CodeMender.

What to watch next

The broader benchmarking for Gemini 4 Carbon is still underway, and Google acknowledges that more advanced internal testing is needed to improve resistance to sophisticated attacks. Watch for updates on how the model performs in real-world drug discovery and how the /swarm feature evolves for developers.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by EdTech Innovation Hub · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news