Gemini 4 Carbon is now available

Google DeepMind has made Gemini 4 Carbon generally available to developers, enterprises, scientists and consumers, following an earlier limited release to cybersecurity defenders through its Fairwind Program two weeks ago. The announcement came from Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. Carbon sits alongside Gemini 4 Argon in Google's frontier lineup; Argon remains the recommended default for most software engineering and knowledge work, while Carbon targets tasks that stress a model's limits, such as complex multi-step coding and debugging and open-ended scientific research. According to the company, Carbon reaches those results by spending more tokens than Argon to explore, reason and run extended verification loops.
Pricing matches Argon's introductory rate: $2 per million input tokens and $10 per million output tokens. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next-best external model. It also placed second on a new Multimodal Knowledge Agent benchmark built with Artificial Analysis, which tests how agents reason over dense files containing charts, images, blueprints, schematics and tables, while coming in at more than 6X lower average cost per task than the top model. On financial research, Google reports Carbon matched the best model tested on Hebbia's E2E Financial Agentic evaluation and ranked first among 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark, which covers legal expertise, agentic tool use and execution, coding and data engineering, and perception and context understanding.
A central part of the release is /swarm, a feature previewed in Google Antigravity that turns a single prompt into a coordinated team of agents tracked in real time. Developers can assign Carbon to roles needing the deepest reasoning, such as architectural design, coordination and auditing, and pair it with faster, lighter models working in parallel isolated workspaces alongside adversarial verifiers that test every change. Demonstrations include a multi-agent team that turned seven public City Council planning PDFs, with 206 extracted engineering drawing views, into an interactive 906K-triangle Three.js reconstruction of Google's Platform 37 office over 10 overnight loops. Other examples include an open-world 3D game exploring the Mountain View campus and a reimagined Google Merchandise Store built in two prompts.
For cybersecurity, Google describes Carbon as its strongest cyber model to date, leading on benchmarks for finding (SEC-bench Pro), exploiting (ExploitGym) and patching (CWE-bench v1) vulnerabilities. In Project Fortify, Google's autonomous scanning, triage and remediation system, Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, and delivered over 25% relative improvement on Argon's recall rate when testing open source repositories. Google's Cloud Vulnerability Research team used Carbon to build a proof of concept showing two bugs in a popular open-source project were exploitable, comparable in severity to a recent exploit the post calls PHP2Shell, prompting maintainers to re-classify the bug as critical and patch it; details await responsible disclosure. Wiz's Scan for Good program used Carbon to find 0-day vulnerabilities in two major SaaS providers in under two hours.
On safety, Google says it strengthened misuse safeguards given the stronger cyber capabilities. In simulated attack environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent, lowering the per-attempt breach rate to 0.8% from 3.1% on Argon, and cut automated cyberattack query success to 2.17% in Gray Swan's testing versus 9.86% on Argon. Trajectory Labs found zero new universal jailbreaks across 132 hours and more than 13,000 requests. On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks versus 70% for Argon, declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time. Google also restates a 0% prompt injection attack success rate on the Gray Swan IPI benchmark while acknowledging that robustness against sophisticated attacks in some distributions could still improve.
The release includes scientific results. Researchers used Carbon within AlphaProof Code to generate millions of lines of Lean code and mathematically prove key security properties of the cryptographic protocol behind Signal in weeks rather than years. Carbon also helped diagnose subtle wet-lab artifacts in enzyme assay development and, across seven low-homolog protein families from the CASP17 blind test set, raised the number of homologs retrieved by 40-fold on average, which could feed structure prediction models like AlphaFold 3 that rely on multiple sequence alignments. Google is extending its scientist access program, which already offered Argon free for six months to 10,000 scientists, to include Carbon for current and future participants, with applications open on a rolling basis.
Availability spans the Gemini API paid tier in AI Studio and Android Studio, Gemini for business, Google Antigravity, CodeMender, and for AI Ultra subscribers the Gemini app and AI Mode in Search. The Fairwind Program remains open to cybersecurity defenders on a rolling basis.
Why it matters: Product teams and developers now have a second frontier option where the trade-off is token spend against depth of reasoning, at identical per-token pricing—so cost per completed task, not per token, becomes the deciding metric. The /swarm preview suggests Google is pushing multi-agent orchestration as a first-class workflow, which may change how teams design coding and research pipelines. Security teams should note the claimed gains in vulnerability discovery, though the strongest cyber capabilities stay gated behind the Fairwind Program rather than open API access.
Based on reporting from the original publisher. Visit the source for full context and later updates.