Google Releases Gemini 4 Carbon: API Introductory Price of $2 per Million Input and $10 per Million Output Tokens Ends November 30, After Which $4 and $20 Apply to the Public-Access Model; the Cyber Results Apply Only Through the Fairwind Program

On November 10, 2026, Google made Gemini 4 Carbon available to developers, enterprises, scientists and consumers. The announcement was posted on Google's blog under the name of Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect. Google calls Carbon its strongest model, with "frontier performance" in cybersecurity and in long-horizon agentic work such as complex knowledge work, software engineering and scientific reasoning. Almost every performance statement in the post is reported by Google, or is a quote from a partner company that Google chose to publish. For the four benchmarks in its main comparisons, CursorBench 4.0, WANDR, AutomationBench and MKA, the text gives rankings and cost ratios but no absolute scores, so this report keeps each claim as a claim, with the comparison Google names next to it.

Carbon sits above Gemini 4 Argon, which Google released earlier. Google says Carbon went first to cybersecurity defenders about two weeks ago through its Fairwind Program. Argon, it says, remains "the ideal choice for the majority of software engineering and knowledge work". Carbon is meant for tasks that test the limits of a model, such as complex multi-step coding and debugging and open-ended scientific research. To do that it uses more tokens than Argon to explore, reason and run longer verification loops. That is a trade-off the reader should note: a model that thinks longer can cost more per request even when the price per token is the same.

On benchmarks, Google says Carbon ranks number one on CursorBench 4.0, on Perplexity's WANDR and on Zapier's AutomationBench, "at more than 2X lower cost per task compared to the next best external model". The post gives no scores and no dollar cost per task for these three benchmarks, so the "2X" is a ratio, reported by Google, against an unnamed competitor. Google calls these "independent benchmarks" but the text does not say who ran the measurements. Google also launched a benchmark with Artificial Analysis, called Multimodal Knowledge Agent (MKA), which tests how a model works through dense files with charts, images, blueprints, schematics and tables. On MKA, Carbon ranks second, and Google says its average cost per task is more than six times lower than that of the top model. The model that ranks first is not named in the text. In pre-release tests, Google says 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.

On price and access, the post is specific. Carbon is available through the API in Google AI Studio and Gemini for business to all paying developers and enterprises. The introductory price is the same as Argon: $2 per million input tokens and $10 per million output tokens. A footnote says this introductory period ends on November 30, 2026. After that date the price of $4 per million input tokens and $20 per million output tokens applies, and the footnote says this covers the Carbon model intended for public access only. Carbon is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. A separate Google developer post adds that the paid API tier is in Google AI Studio and Android Studio, and that Ultra subscribers can use it in AI Studio and Antigravity. "Rolling out" means not every user has it yet. Google does not give a date by which all Ultra users will have it.

GitHub announced the same day that Carbon is generally available in GitHub Copilot. It is open to Copilot Pro+, Max, Business and Enterprise users, and the model picker lists it in Visual Studio Code, Visual Studio, the Copilot CLI, the Copilot coding agent, the Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. GitHub says the model is billed at the provider's list price under usage-based billing, and that the rollout is gradual. On Copilot Business and Enterprise plans, administrators manage access through the model policy. Under default model enablement, new models are switched on unless an administrator has turned off the global default or disabled this model.

For large projects, Google is previewing a feature called /swarm in Google Antigravity. It turns one prompt into a team of agents that the user can track in real time. A developer can give Carbon the roles that need the deepest reasoning, such as architecture, coordination and auditing, and pair it with faster, lighter models in parallel isolated workspaces with adversarial verifiers. Google shows several demonstrations. In one, seven public City Council planning PDFs containing 206 extracted engineering drawing views and a single prompt led a team of Carbon agents to build an interactive 906K-triangle Three.js reconstruction of Google's Platform 37 office over 10 overnight loops. These are Google's own showcase projects. The post does not say how many attempts were made, or how many failed.

The post also collects customer statements. HubX says that across 16 real bug fixes in its 14,000-test codebase, Carbon achieved a 94% resolution rate with zero regressions. With 16 fixes as the base, one fix is about six percentage points, so the figure is a small sample. Box says Carbon scored 100% on public sector tasks in its testing, without giving the number of tasks. OpenCode says Carbon rebuilt its desktop app overnight with a new Rust UI library. Cartwheel says it spent six days with minimal steering on a cloth simulation. These are quotes from companies that chose to speak at launch. They are not independent measurements.

The cybersecurity part has a condition that is easy to miss. The results come from "Carbon for cyber", which is reached through the Fairwind Program. Through that program, Google says Carbon leads on SEC-bench Pro for finding vulnerabilities, ExploitGym for exploiting them and CWE-bench v1 for patching them. Google gives no scores for these three benchmarks. It does give figures from an internal benchmark: in Project Fortify, Google's system for scanning, triage and remediation, it says Carbon hit 100% recall on scanning and 95% precision on triage on its internal benchmark. Partners add their own accounts: SentinelOne says Carbon for cyber found four additional verified vulnerabilities in a controlled kernel driver evaluation, and Armadin says it reached results in half the steps of its predecessor. The post says cybersecurity defenders can apply to join the Fairwind Program and are admitted on a rolling basis, but it does not list the full eligibility or review criteria. A general developer should therefore not assume the cyber results apply to the model they can buy.

On science, Google says Axiom Bio, a drug toxicity research company, found Carbon to be the leading model for translational science and is moving its production analyses and agents to both Carbon and Argon. Four weeks ago Google opened applications for a program to give 10,000 scientists free access to Argon in Antigravity for six months. It now expands that access to Carbon for current and future participants. Onboarding continues on a weekly rolling basis, subject to eligibility, capacity and selection. Google also says Carbon builds on Argon's safety and security stack, with added protections against misuse and prompt injection, and points to a separate Safety and Security report and a model card. The post itself lists some safety test figures. For example, Google says Carbon passed 92% of tasks on its own 375-task agentic honesty benchmark, against 70% for Argon; some other figures come from named outside testers such as Gray Swan and Trajectory Labs. This report did not reproduce any of them, and it did not read the separate report or the model card, so it makes no statement about their findings.

What the reader can check today is limited. The price, the end date of the introductory rate, the list of products and the Copilot plans are stated facts. The rankings and cost ratios are Google's account of results on benchmarks it calls independent, and the post does not say who ran those measurements. The customer results are the customers' own accounts, chosen for the launch post. This report reproduced none of them. Charts and videos in the post were not read for this report. Before choosing Carbon over Argon for a workload, a team should run its own tests, measure tokens per task as well as price per token, and plan for the new price that applies after the introductory period ends on November 30, 2026.