Gemini 4 Carbon Launches: Google's Strongest AI Model Explained

Gemini 4 Carbon is officially available. Google made its new top-tier "strongest model" generally available on November 10, 2026, positioning it above the recently launched Gemini 4 Argon for the hardest agentic, scientific, and cybersecurity work — while keeping Argon as the recommended default for most everyday coding and knowledge tasks. Carbon ranks #1 on three independent agentic benchmarks at more than double the cost-efficiency of the next-best outside model, and Google is backing it with firm pricing and a wider rollout across its developer and consumer products.

Here is what changed with general availability, how Carbon stacks up against Argon and Gemini 4 Hydrogen, the fast model already powering Search, and exactly what it costs and where you can use it.

What is Gemini 4 Carbon?

Gemini 4 Carbon is the new top-tier model in Google's Gemini 4 lineup, which Google DeepMind SVP Koray Kavukcuoglu describes as the company's strongest model, built for frontier performance in cybersecurity and in long-horizon agentic work spanning knowledge work, software engineering, and scientific reasoning. It became generally available on November 10, 2026, roughly two weeks after Google first released it privately to cybersecurity defenders through its Fairwind Program.

Carbon is explicitly not a replacement for Argon. Google's own guidance is that Argon remains the better choice for most software engineering and knowledge work, because it is faster and cheaper to run on routine requests. Carbon is aimed at harder problems — complex multi-step coding, debugging across large codebases, and open-ended research — and it uses more tokens than Argon to explore multiple approaches, reason through them, and run its own verification loops before answering. That extra deliberation is also why Carbon costs more to run in practice even though its headline per-token price matches Argon's.

How Carbon differs from Argon and Hydrogen

Google now has three distinct Gemini 4 model tiers in production, and the naming can be confusing — see this breakdown of every Gemini 4 variant for the full lineup. In short:

• Gemini 4 Hydrogen is the fast, low-latency model, already powering Google Search and consumer-facing products where speed matters more than depth.
• Gemini 4 Argon is the previous flagship and Google's recommended default for most software engineering and general knowledge work — strong, fast, and the better value for routine tasks.
• Gemini 4 Carbon is the new strongest model, reserved for the hardest long-horizon agentic tasks, deep research, and cybersecurity work, where it burns more tokens exploring and checking its own work in exchange for materially better results.

Google's framing is a deliberate trade-off: Carbon is not meant to replace Argon for day-to-day use. It is meant to be reached for when a task is genuinely hard enough to justify the extra reasoning and expense.

Benchmark results: where Carbon leads

Google's announcement leans heavily on third-party agentic benchmarks rather than Google's own internal scoring. According to the company:

• Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench — three independent agentic-workflow benchmarks — at more than 2X lower cost per task than the next-best external model.
• Carbon ranks #2 on the Multimodal Knowledge Agent (MKA) benchmark, built with Artificial Analysis, while costing more than 6X less on average per task than the top-performing model on that benchmark.
• In a pre-release evaluation with Hebbia, Carbon matched the best model tested on an end-to-end financial agentic research task.
• On Balyasny Asset Management's General Financial Intelligence benchmark, Carbon ranked at the top of 23 frontier model configurations tested.

Google also highlights cybersecurity-specific results gathered through the Fairwind Program, where Carbon leads on SEC-bench Pro (finding vulnerabilities), ExploitGym (exploiting them in controlled settings), and CWE-bench v1 (patching them). In Google's internal Project Fortify testing, Carbon achieved 100% recall on scanning and 95% precision on triage, with recall on real open-source repositories improving more than 25% relative to Argon.

Safety and security results

Google frames Carbon's safety profile as a direct improvement over Argon, backed by a published Gemini 4 model safety and security report. In simulated cyberattack environments (ExploitBench and CyScenarioBench) run against an adaptive red-teaming agent, Carbon blocked 99% of attack trajectories. Its per-attempt breach rate fell to 0.8%, down from 3.1% on Argon. Against Gray Swan's automated cyberattack queries, Carbon's success rate for attackers was 2.17%, compared with 9.86% on Argon. External testing by Trajectory Labs found zero new universal jailbreaks across 132 hours and more than 13,000 requests, and Carbon reportedly achieved a 0% attack success rate on Gray Swan's indirect prompt injection benchmark — a result Google says it announced two weeks before the GA launch, while cautioning that robustness against sophisticated attacks can still improve.

On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, compared with 70% for Argon. Google says Carbon declined to fabricate missing experimental data 98% of the time and flagged methodological flaws in research tasks 68% of the time — both framed as evidence that Carbon is more reliable to delegate open-ended, unsupervised work to.

Those safety gains are also the backbone of Carbon's cybersecurity pitch: Google cites a Wiz "Scan for Good" exercise in which Carbon found zero-day vulnerabilities in two major SaaS providers in under two hours, and a Google Cloud Vulnerability Research effort that used Carbon to prove two open-source bugs were exploitable, leading both to be reclassified as critical and patched.

Pricing: what Gemini 4 Carbon costs

Carbon launches at the same headline API pricing as Argon, per Google's own Gemini API pricing documentation: $2 per million input tokens and $10 per million output tokens through November 30, 2026. Starting December 1, 2026, those rates double to $4 per million input tokens and $20 per million output tokens for the public-access model. Neither Carbon nor Argon is available on a free tier.

Because the per-token price is identical to Argon's, the real cost difference between the two models in production comes down to usage: Carbon's longer reasoning and self-verification loops on hard tasks mean it typically consumes more tokens per completed task, even though neither model's rate card favors one over the other.

Where Gemini 4 Carbon is available

Google is rolling Carbon out across its developer, enterprise, and consumer surfaces rather than restricting it to a single product:

• Developers: the paid tier of Google AI Studio and Android Studio, with a published Gemini 4 Carbon developer guide and migration documentation.
• Enterprises: Gemini for business, accessible via API or model dropdown, plus Google Antigravity and CodeMender.
• Google AI Ultra subscribers: a rolling-out release in Google Antigravity, the Gemini app, and AI Mode in Search.
• Cybersecurity defenders: continued access through the Fairwind Program, which first received Carbon roughly two weeks before this GA launch.
• Scientists and researchers: free access to both Argon and Carbon through Antigravity's highest usage tier for six months, granted on a rolling basis subject to eligibility and capacity, via an application program Google says has already onboarded thousands of the 10,000 scientists it first invited.

Google is also previewing a feature called /swarm inside Google Antigravity, which turns a single prompt into a coordinated team of agents: Carbon takes roles that need deep reasoning — architecture, coordination, and auditing — while faster, lighter models handle parallelizable work and adversarial verifiers check the results.

The Fairwind Program and early access

Unlike a typical model launch, Carbon's rollout had a staged, audience-specific start. Google first gave cybersecurity defenders early access through the Fairwind Program about two weeks before the November 10 GA date, using that window to generate the vulnerability-scanning and patching benchmark results Google now cites publicly. Separately, Google's scientist access program — which opened applications about four weeks before this announcement and has already onboarded thousands of an initial 10,000 invited scientists for free six-month access to Argon — is now being expanded to include Carbon as well, with Google saying future frontier models will be added to the same program as they launch. Applications for both programs remain open.

This two-track approach — specialists and researchers first, broad commercial and consumer availability second — mirrors how Google introduced Argon, and it is becoming the company's standard playbook for frontier model releases: see what was confirmed about Carbon before this launch for how closely the pre-release signals tracked today's announcement.

Early use cases Google is showcasing

Alongside the technical benchmarks, Google's announcement leans on a set of concrete demonstrations meant to show Carbon handling long, unsupervised work. These include Carbon reconstructing a 906,000-triangle 3D model of one of Google's own office buildings from seven public city council PDFs containing 206 drawing views, run across ten overnight work loops with no human in the loop; an interactive Three.js exhibition recreating the Eiffel Tower's 1887–1889 construction; and a redesign of the Google Merchandise Store with five distinct art-direction concepts and new product imagery.

On the scientific side, Google cites Axiom Bio using Carbon as its leading model for translational science research, where it says the model identified new toxicity mechanisms, and a protein-structure prediction test across seven low-homolog protein families from the CASP17 blind test set, where Carbon increased retrieved homologous structures roughly 40-fold on average without any additional retraining. Google DeepMind's AlphaProof Code effort also used Carbon to generate formal Lean proofs of security properties in Signal's cryptographic protocol.

What happens next

The near-term date that matters most for anyone building on Carbon is November 30, 2026 — the last day of introductory pricing before per-token rates double on both input and output. Enterprises and developers evaluating Carbon for production agentic workloads have roughly two and a half weeks from this GA announcement to budget around that change. Beyond pricing, watch for the wider Ultra-subscriber rollout of Carbon inside the Gemini app and AI Mode in Search to reach general availability (Google describes it as still "rolling out" rather than fully live), and for results from the /swarm multi-agent preview in Google Antigravity, which is the clearest signal yet of how Google expects Carbon to be used day to day: not as a single chat model, but as the senior reasoning layer coordinating a team of cheaper, faster models.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage