Google has made Gemini 4 Carbon available today to developers, enterprises, scientists, and consumers. According to the company, Carbon is the most capable Gemini model to date, with frontier performance in cybersecurity and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning.
Two weeks earlier Carbon was released to a set of cybersecurity defenders via the Fairwind Program. Following the recent launch of Gemini 4 Argon, Carbon is positioned to push capability further: Argon remains the preferred choice for most software engineering and knowledge-work tasks, while Carbon targets tasks that push a model’s intelligence limits—complex multi-step coding and debugging, open-ended scientific research—using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.
Performance and production economics
Google cites independent benchmarks showing Carbon pairs frontier accuracy with practical production economics. The company reports Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, and delivers more than 2x lower cost per task versus the next best external model.
To apply Carbon to larger, complex projects, Google is previewing the /swarm feature in Google Antigravity, which turns a single prompt into a coordinated team of agents tracked in real time. Developers can assign Carbon to roles requiring the deepest reasoning—architectural design, coordination, auditing—while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change.
Availability and pricing
Gemini 4 Carbon is available via the API in Google AI Studio and Gemini for business to all paying developers and enterprises at the same introductory price as Gemini 4 Argon: $2 per 1M input tokens and $10 per 1M output tokens. It is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search.
Google states that after the introductory period ends on November 30, 2026, a higher price will apply for public access to Carbon: $4 per 1M input tokens and $20 per 1M output tokens.
Knowledge work, document processing and multimodal agents
Carbon combines visual understanding, reasoning, and exhaustive retrieval to synthesize grounded insights across large sets of dense documents and data in legal, finance, and enterprise operations. The model follows strict schemas and formatting to generate precise spreadsheets, presentations, and reports.
Because existing benchmarks do not fully capture how agents reason over dense multimodal files (charts, images, blueprints, schematics, tables), Google partnered with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark to test multimodal agentic performance. Carbon ranks #2 on MKA while claiming more than 6x lower average cost per task versus the top-performing model.
Google shares examples of Antigravity /swarm demonstrations: Carbon agent teams transformed historical photos and engineering plates into an interactive Three.js 3D exhibition of the Eiffel Tower’s 1887–1889 construction from a single prompt. Using seven public City Council planning PDFs (206 extracted engineering drawing views) and a single prompt, a multi-agent Carbon team built an interactive 906K-triangle Three.js reconstruction of Google’s Platform 37 office across 10 overnight loops. Other demos include a multiscale textile visualizer based on an academic paper and a pipeline that converted a cinematic underwater Atlantis scene from Blender into a real-time Three.js browser experience.
Cybersecurity capabilities
Gemini 4 Carbon for cyber is presented as Google’s strongest cybersecurity capability so far, with leading benchmark results on finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities. Carbon inspects software, traces how flaws chain across systems, and generates verified root-cause patches when accessed through the Fairwind Program.
Early internal testing results cited by Google include Project Fortify, an autonomous scanning, triage, and remediation system where Carbon achieved 100% recall on scanning and 95% precision on triage in Google’s internal benchmark. On open-source repository testing Carbon shows over 25% relative improvement on Argon’s recall rate, surfacing previously undetected real-world logic flaws.
Google’s Cloud Vulnerability Research (CVR) team used Carbon to demonstrate that two bugs in a popular open-source project were exploitable vulnerabilities and helped prompt maintainers to re-classify the bug as critical and release a patch; details will be shared after the responsible disclosure process completes. As part of Wiz’s Scan for Good, Carbon helped identify 0-day vulnerabilities in major SaaS providers in under two hours in some cases, compressing weeks of vulnerability research into an afternoon and enabling rapid remediation with affected organizations.
Scientific and mathematical reasoning
Google reports improved performance from Carbon across physics, math, and life sciences: synthesizing experimental data, testing hypotheses, and solving multi-step domain problems. For Axiom Bio, a drug toxicity research company, Carbon is now the leading model for translational science work and has found novel mechanisms of toxicity previously missed; Axiom plans to move production scientific analyses and agents to Carbon and Argon.
Researchers inside Google have used Carbon in Antigravity for several advanced tasks:
- Code verification: An agentic formal verification framework called AlphaProof Code used Carbon (alongside other models) to translate security goals into mathematical theorems and generate millions of lines of Lean code to prove security properties of Signal’s underlying cryptographic protocol in weeks—a task that historically takes expert teams years.
- Enzyme assay development: Given only lab notebooks and raw data, Carbon surfaced hard-to-interpret anomalies, diagnosed experimental artifacts (e.g., incomplete automated mixing), improved chromatography deconvolution, and detected trace sample carryover, accelerating assay development.
- Protein structure prediction support: In a multi-agent setup Carbon autonomously discovered remote homologs across petabase-scale unannotated sequence archives and structure databases. For seven low-homolog protein families from the CASP17 blind test set, Carbon increased the number of retrieved homologs by an average of 40-fold, improving inputs for structure-prediction models without retraining.
Four weeks ago Google opened applications to give 10,000 scientists access to Argon in Google Antigravity free for six months; thousands have already onboarded. Today the company is expanding access to Carbon for current and future participants; applications remain open and onboarding continues weekly, subject to eligibility, capacity, and selection.
Safety, misuse mitigation and alignment progress
Google says it continues to strengthen safety and security mitigations, agentic misuse prevention, model alignment, and harm reduction proportionally with model capability. Because Carbon introduces stronger cyber capabilities, Google built on Argon’s safety stack to harden robustness to misuse and prompt injection.
In simulated cyberattack testing seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from Google’s adaptive red-teaming agent, lowering the per-attempt breach rate to 0.8% (down from 3.1% on Argon). Carbon’s safeguards also reduced the automated cyberattack query success rate to 2.17% in Gray Swan testing versus 9.86% on Argon. External testing by Trajectory Labs reported zero new universal jailbreaks across 132 hours and over 13,000 requests.
On alignment metrics, Google reports Carbon passed 92% of tasks on its 375-task agentic honesty benchmark (Argon scored 70%): it declined to fill missing experimental data 98% of the time and flagged methodological flaws 68% of the time.
Google previously announced Carbon achieved a 0% prompt-injection attack success rate on the Gray Swan IPI benchmark. The company notes that prompt-injection robustness remains an area for continued improvement and that advanced automated red teaming continues to identify areas for further strengthening.
More details are available in the Gemini 4 Carbon Safety and Security report.
How to get started
Starting today, Carbon joins Argon across Google products:
- For developers: build with the Gemini API paid tier in AI Studio and Android Studio. Google AI Ultra subscribers can try it in Google Antigravity and AI Studio.
- For enterprises: try it in Gemini for business via the API or model dropdown, Google Antigravity, or CodeMender.
- For everyday users: Carbon is available in the Gemini app and AI Mode in Search for Google AI Ultra subscribers.
- For scientists & researchers: apply for access to Argon and Carbon via Google Antigravity’s highest usage tier free for six months (onboarding is rolling and subject to eligibility and capacity).
- For cybersecurity defenders: apply to join the Fairwind Program for rolling access to Carbon for cyber.
Google points readers to the model card, the updated Gemini 4 developer guide, and documentation for best practices and migration guidance.
(Key pricing points reiterated: introductory price $2/1M input and $10/1M output; after introductory period ends on 2026-11-30 public access pricing becomes $4/1M input and $20/1M output.)
