Glonce

Google DeepMind releases Gemini 4 Carbon at Argon's API price

modelsThe Glonce editorsGoogle DeepMind🔥80Glonce rating80Impact85Novelty70Relevance100Surprise45Players95Credibility88
Google DeepMind releases Gemini 4 Carbon at Argon's API price

Google DeepMind has made Gemini 4 Carbon available to developers, enterprises, scientists and consumers. Google calls Carbon its strongest model, with frontier performance in cybersecurity and in long-horizon agentic workflows across complex knowledge work, software engineering and scientific reasoning. Two weeks earlier Carbon went to a set of cybersecurity defenders through the Fairwind Program.

Carbon follows the recent launch of Gemini 4 Argon, and Google does not present it as a replacement. Argon remains, in Google's words, the ideal choice for most software engineering and knowledge work. Carbon is meant for tasks that test the limits of a model's intelligence, such as complex multi-step coding and debugging and open-ended scientific research, and it uses more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops.

Availability and price: Carbon is offered through 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 million input tokens and $10 per million output tokens. It is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. Google is also previewing /swarm in Google Antigravity, which turns one prompt into a coordinated team of agents that can be tracked in real time. Developers can give Carbon the roles that need the deepest reasoning, such as architectural design, coordination and auditing, and pair it with faster, lighter models in parallel isolated workspaces with adversarial verifiers that test every change.

Benchmarks: Google says Carbon ranks #1 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. With Artificial Analysis it launched a new Multimodal Knowledge Agent (MKA) benchmark, which tests agents on dense files with charts, images, blueprints, schematics and tables; Carbon ranks #2 there at more than 6X lower average cost per task than the top performing model. In pre-release evaluations Carbon matched the best model tested on Hebbia's E2E Financial Agentic evaluation and ranked at the top out of 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark. Google also says Carbon builds and verifies full web apps, checks the rendered UI, and does better on 3D game development and creative coding.

Cybersecurity: accessed through the Fairwind Program, Carbon leads on key benchmarks for finding (SEC-bench Pro), exploiting (ExploitGym) and patching (CWE-bench v1) vulnerabilities. In Project Fortify, Google's autonomous system for scanning, triage and remediation, Carbon hit 100% recall on scanning and 95% precision on triage on Google's internal benchmark, and showed an over 25% relative improvement on Argon's recall on open source code repository testing. Google's Cloud Vulnerability Research team used Carbon to prove that two bugs in a popular open-source project were exploitable, comparable in severity to the recent PHP2Shell exploit; maintainers re-classified the bug as critical and issued a patch, and details will follow after responsible disclosure. In Wiz's Scan for Good program, Carbon uncovered 0-day vulnerabilities at two major SaaS providers in under two hours, which Google describes as compressing weeks of research into one afternoon.

Science and mathematics: drug toxicity research company Axiom Bio calls Carbon the leading model for translational science; it found toxicity mechanisms the company had not found before, and Axiom is moving production analyses and agents to Carbon and Argon. Google researchers used Carbon in Antigravity for three projects. First, AlphaProof Code, a new agentic framework for formal verification, used Carbon to generate millions of lines of Lean and prove key security properties of the Signal protocol code in weeks, a task Google says historically takes expert teams years. Second, given only lab notebooks and raw data, Carbon spotted data anomalies and used them to diagnose wet lab artifacts such as incomplete automated mixing in deep wells, improved methods for deconvoluting chromatography peaks, and caught trace sample carryover. Third, in protein structure prediction, a multi-agent setup with Carbon searched sequence archives and structure databases for remote homologs; across 7 low-homolog protein families from the CASP17 blind test set it increased the number of homologs retrieved by 40-fold on average, without retraining the structure models. Google adds that broader benchmarking is still underway.

Scientist access: four weeks earlier Google opened applications to give 10,000 scientists free access to Argon in Antigravity for six months, and thousands are onboarded. Access is now extended to Carbon for current and future participants, and the program will keep offering the latest frontier models. Applications remain open, with weekly rolling onboarding subject to eligibility, capacity and selection.

Safety and security: in simulated cyberattack environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent, and the per-attempt breach rate fell to 0.8% from 3.1% on Argon. In Gray Swan's testing the automated cyberattack query success rate was 2.17% against 9.86% on Argon. Trajectory Labs found zero new universal jailbreaks across 132 hours and over 13,000 requests. On Google's 375-task agentic honesty benchmark Carbon passed 92% of tasks, against 70% for Argon, declining to fill in missing experimental data 98% of the time and flagging methodological flaws 68% of the time. Two weeks earlier Google reported a 0% prompt injection attack success rate on the Gray Swan IPI benchmark, but it says prompt injection robustness remains an area for continued improvement. A separate Gemini 4 Carbon Safety and Security report has the details.

Key facts

  • Google DeepMind calls Gemini 4 Carbon its strongest model; API price is the same introductory $2 per million input tokens and $10 per million output tokens as Gemini 4 Argon.
  • Carbon is available via the API in Google AI Studio and Gemini for business to paying developers and enterprises, and is rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.
  • Google says Carbon ranks #1 on CursorBench 4.0, WANDR and AutomationBench at more than 2X lower cost per task than the next best external model, and #2 on the new MKA benchmark at more than 6X lower average cost per task than the top model.
  • In simulated cyberattack tests the per-attempt breach rate fell to 0.8% from 3.1% on Argon, and Carbon passed 92% of tasks on Google's 375-task honesty benchmark versus 70% for Argon.
  • Argon stays Google's pick for most software engineering and knowledge work; Carbon is for the hardest tasks and uses more tokens.

Why it matters

This is Google's top-tier model reaching general use only after a short staged rollout, first to cybersecurity defenders and now to paying developers, enterprises, scientists and consumers. It sits above Gemini 4 Argon, launched recently, at the same introductory token price. Google's pitch is accuracy together with cost: top ranks on three third-party agent benchmarks at more than 2X lower cost per task than the next best external model. It also puts cyber capability at the centre of the release, with safety numbers published alongside.

Who it affects

Developers and enterprises on the paid Gemini API tier in Google AI Studio and Gemini for business, and Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Security teams are a stated focus, through Fairwind, Project Fortify and the Wiz Scan for Good work. Scientists in the free six-month Antigravity program get Carbon access too. Teams in legal, finance and enterprise operations are named as targets for the document-heavy knowledge work.

How to use it

Developers can call Carbon through the API in Google AI Studio or Gemini for business, at $2 per million input tokens and $10 per million output tokens, the introductory price shared with Argon. Ultra subscribers get it in Antigravity, the Gemini app and AI Mode in Search as the rollout proceeds. Google advises Argon for most software engineering and knowledge work and Carbon for the hardest multi-step coding, debugging and research tasks, bearing in mind that Carbon spends more tokens. In Antigravity, /swarm is a preview that lets Carbon take roles like architecture, coordination and auditing while lighter models work in parallel. Scientists can apply to the free program, which has weekly rolling onboarding subject to eligibility, capacity and selection.

How solid is it

This is a first-party launch post, so every number is Google's own reporting. Some results come from outside parties: the ranks on CursorBench 4.0, WANDR and AutomationBench, the MKA benchmark built with Artificial Analysis, Gray Swan's testing and Trajectory Labs' jailbreak testing. The Fortify figures (100% recall, 95% precision) are on Google's internal benchmark, and the 375-task honesty benchmark is Google's too. The post gives ranks and cost ratios for the third-party benchmarks, not scores, and does not name the next best external model or the top model on MKA. The Balyasny and Hebbia results are described as pre-release evaluations. The Signal verification result is Google's own account.

Risks and caveats

The price is called introductory, and the visible text does not say how long it lasts or what comes after. Carbon uses more tokens than Argon, so cost per task may differ from the per-token price. The cyber results depend on access through the Fairwind Program. Google itself says prompt injection robustness remains an area for continued improvement and that its automated red teaming shows models could be improved against sophisticated attacks in some distributions. The protein result is hedged, since broader benchmarking is still underway. Details of the open-source bugs found by Google's CVR team will come only after responsible disclosure. The rollout to Ultra subscribers is still in progress.

“Carbon is our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning.”

— Google DeepMind, Gemini 4 Carbon announcement