Google on November 10, 2026 began rolling out Gemini 4 Carbon, which it describes as its strongest model, to Google AI Ultra subscribers in AI Mode in Search, the Gemini app and Google Antigravity, and made it available through the Gemini API 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. The announcement, written by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, says Carbon uses more tokens than Argon to reason through hard problems, and it names no route to the model for subscribers on the cheaper AI Pro plan.

In Short

Google released what it calls its strongest AI model, Gemini 4 Carbon, to paying developers and businesses on November 10, and started putting it into AI Mode in Search and the Gemini app for people on its top Google AI Ultra plan. It matters to marketers because Carbon now writes some Search answers and can power the reporting, analysis and coding tools agencies build on Google's platform, at the same price per unit as the previous model even though Google says it does more work per job. Nearly every performance figure comes from Google or partners it chose, and the post says nothing about ads, citations or Google Ads.

What Google released on November 10

The post, titled "Gemini 4 Carbon is now available" and dated November 10, 2026, sits in the Gemini Models section of Google's blog and carries a 12-minute reading label. Its opening paragraph sets out the claim: "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."

Carbon is not new to PPC Land readers. When Alphabet reported third-quarter results on October 28, chief executive Sundar Pichai introduced Carbon as Google's strongest frontier model, said it ranked first on the Artificial Analysis Intelligence Index with a score of 59, and said it would go first to a subset of trusted cyber defenders while being tested by the US government. The November 10 post confirms the first part: "Two weeks ago, we released Carbon to a set of cybersecurity defenders via our Fairwind Program." It does not mention the Intelligence Index, and it says nothing about the government testing Pichai described.

A week before that, in an interview published on October 21, Pichai had been asked whether rumours of a "carbon" model pointed to a coming release, and said only that Google always runs internal versions of its frontier models under code names. Twenty days later, the code name is a product.

Google does not present Carbon as a replacement for Argon, the first Gemini 4 model, released on October 13. According to the post, "Argon remains the ideal choice for the majority of software engineering and knowledge work, while Carbon is built for tasks that test the limits of a model's intelligence like complex, multi-step coding and debugging and open-ended scientific research and discovery." Two Gemini 4 models now sit side by side, priced identically and separated, in Google's description, by how hard the task is and how long the model is allowed to work on it.

Same rate as Argon, more tokens per task

The pricing sentence is short. According to Google, Carbon is available 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. The word "introductory" carries a footnote, and the body of the post does not say how long the rate lasts or what replaces it.

Those are the figures Google attached to Argon four weeks earlier, when it priced Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, for paid users only. Google's frontier input rate has not moved since November 2025, when it priced Gemini 3 Pro in preview at $2 per million input tokens and $12 per million output tokens for shorter prompts.

A matching rate card does not mean a matching bill. The same paragraph that positions Carbon above Argon says it gets there by "using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." Every additional token of reasoning is billed at the same $10 per million output rate. On an identical job, then, Carbon would cost more than Argon wherever it generates more output - a PPC Land reading of Google's own wording. How much more? The post gives no ratio between the two models' token use and no cost-per-task comparison between them.

That gap matters because inference cost is where model spending accumulates once a tool is in daily use. Agencies were already metering AI media-buying agents by the token by September, and on October 8 Google Cloud added project spend caps that pause Gemini agents at the limit. A model marketed on working longer runs straight into both controls.

Cost per task, as Google measures it

Google frames Carbon's economics in a different unit. "Across independent benchmarks designed to test model performance in real workflows, Carbon pairs frontier accuracy with practical production economics," the post says. It "ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external model."

Two words in that sentence do a lot of work. "External" excludes Google's own models, so Argon is not part of the comparison. "Next best" is not named. The post does not identify which competing model, from which company, sat second on each of the three benchmarks, or what each benchmark run cost.

Cost per task has become the metric of choice across the industry. Anthropic's IPO filing claimed that Claude Opus 5.5 cuts cost per task by 40%, with no definition of a task in the summary. Google's figure has the same limitation. A lower cost per completed task can coexist with a higher monthly bill when cheaper tasks lead teams to run more of them.

Capacity is the other constraint. On October 21, Pichai said Argon had processed almost 25 trillion tokens in its first 72 hours and that Google remained short of compute. Carbon, by Google's own description, consumes more per task than the model that strained that capacity. The post gives no usage limits for API customers or subscribers.

Who can reach Carbon

Google's "Getting started" section lists five routes:

  • Developers can build with the Gemini API paid tier in AI Studio and Android Studio. Google AI Ultra subscribers can also try it in Google Antigravity and AI Studio.
  • Enterprises can use it in Gemini for business through the API or a model drop-down, in Google Antigravity, or in CodeMender, Google's code security agent.
  • Everyday users can try Carbon in the Gemini app and AI Mode in Search "if you're a Google AI Ultra subscriber."
  • Scientists and researchers can apply for both Argon and Carbon through Google Antigravity's highest usage tier, free of charge for six months, "subject to eligibility, capacity, and selection."
  • Cybersecurity defenders can apply to the Fairwind Program for access to "Carbon for cyber" on a rolling basis.

The consumer route is the narrowest Google has drawn for a Gemini 4 model. Argon reached AI Mode and the Gemini app for both AI Pro and Ultra subscribers, with Pro priced at $19.99 a month. Carbon's post names Ultra alone, a plan PPC Land has reported starting at $99.99 a month. On those prices, the entry point for Carbon in AI Mode is five times the subscription that was enough for Argon - a PPC Land calculation, since the post itself quotes no consumer prices.

Free users are further back still. From October 9, free Gemini accounts kept only Flash-Lite, losing access to Flash and Pro. Nothing in the November 10 post points to free access to Carbon, and it does not say whether Ultra subscribers keep Argon as an option alongside it.

A model in AI Mode, and a silence on ads

For search marketers, the most consequential line in the post is the shortest: Carbon "is also rolling out to all Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search."

That makes Carbon the third model change in AI Mode in a little over two months. AI Mode moved to Gemini 3.8 Flash on September 2, took on Argon for Pro and Ultra subscribers on October 13, and now adds Carbon for Ultra. Each change alters the model writing answers above, around and sometimes instead of the links that send visitors to websites.

The post contains no sentence about advertising. It does not say whether answers written by Carbon carry ads in the same positions as other AI Mode answers, whether Carbon cites more or fewer sources than Argon, how many links it shows, or in which countries and languages it is available. Argon's post had described a global rollout; Carbon's does not use the word.

Those gaps sit against an advertising product that is changing quickly. On October 28, Google said it was taking ads in AI Mode to 10 more countries, from Australia and Canada to Nigeria and Singapore, with no separate opt-out for advertisers. In early November, searchers documented up to four ads beneath a single AI Mode answer. And Alphabet's own results showed the commercial tension: the company said AI Mode's query growth weighed on search monetisation, and its Class A shares fell 12.6% on October 29. The same week, Reddit chief executive Steve Huffman blamed Google's AI search features for falling referrals as Reddit warned of a drop in daily users.

Carbon in AI Mode is, for now, limited to subscribers on Google's most expensive consumer plan, and Google has not disclosed how many people that is. It is also absent from Google's advertising products in this announcement. The post does not mention Google Ads, Performance Max, Merchant Center or any ad tool.

Spreadsheets, decks and financial models

The section of the post closest to day-to-day agency work concerns documents. Across legal, finance and enterprise operations, Google says, Carbon "combines frontier visual understanding, reasoning, and exhaustive retrieval to synthesize grounded insights across large sets of dense documents and data. It follows strict schemas and formatting to generate polished and precise spreadsheets, presentations, and reports."

Client reporting, media plans and pitch decks are built in exactly those formats. Whether Carbon's output holds up on a marketer's own data is not something the post addresses; its evidence comes from finance.

Google cites three evaluations. On the new Multimodal Knowledge Agent (MKA) benchmark, built with Artificial Analysis to test how agents reason over files containing charts, images, blueprints, schematics and tables, Carbon ranks second, "while being more than 6X lower in average cost per task compared to the top performing model." The post does not name the model in first place. That a model Pichai described as first on one Artificial Analysis index ranks second on another, newer one is a detail the post leaves without comment.

In pre-release testing, according to Google, Carbon "matched the performance of the best model tested" on Hebbia's E2E Financial Agentic evaluation, and ranked at the top of 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark, which spans legal expertise, agentic tool use, coding and data engineering, and perception. The post gives no link to the methods or scores behind either evaluation.

Shanks Salwan, Lead Applied AI Engineer at Hebbia, is quoted in the post: "Gemini 4 Carbon is the strongest model we've evaluated for deep financial research, building cleaner, better-structured Excel models and producing sharper PowerPoint decks. It is a real step forward for diligence workflows." Audrey Liu, Senior AI Engineer at Balyasny, said the firm's benchmark is designed to "evaluate the work financial teams actually do," and that Carbon stood out on tasks combining complex context with structured deliverables such as charts, code and tables.

/swarm: many agents from one prompt

Alongside the model, Google is previewing a feature called /swarm in Google Antigravity, its agentic development environment. According to the post, it turns "a single prompt into a coordinated team of agents you can track in real time." Developers can assign Carbon "to roles that demand the deepest reasoning, like architectural design, coordination, and auditing, while pairing it with faster, lighter models across parallel isolated workspaces and adversarial verifiers that test every change."

The split fits the cost profile described above: Carbon handles planning and checking, while cheaper models do the bulk of the work. How the tokens split between them, and what a typical /swarm run costs, is not disclosed.

Google's demonstrations are visual. In one, a team of Carbon agents turned historical photos and engineering plates into an interactive Three.js 3D exhibition of the Eiffel Tower's 1887-1889 construction from a single prompt. In another, Carbon built an open-world 3D game in which a player explores Google's Mountain View campus as a pelican riding a GBike, coordinating subagents that modelled 3D assets in Blender from Google's Omni and Nano Banana models.

On engineering work, Google says Carbon "performs exceptionally well" on multi-package migrations, cross-subsystem root-cause investigations and complex architectural redesigns, and that for web apps it "runs build-and-verify loops, recovers from errors, and checks both its code and the rendered UI." Two customers are quoted. Kit Langton, software engineer at OpenCode, said: "When we struggled to get our app looking consistent with an existing UI framework, we gave Carbon a list of requirements before closing our day. Overnight, Carbon had built a new Rust UI library from scratch and rewritten our OpenCode desktop app with working demos. It's an incredibly brainy model and now a preferred planning model for our hardest tasks."

Andrew Carr, chief scientist and co-founder of Cartwheel, called Carbon "by far the best frontier model for difficult cloth simulation." His team, he said, "spent 6 days working with it with minimal steering," after which the model had improved an existing soft-body simulation in both speed and fidelity.

Cybersecurity: the claims attach to the Fairwind version

The cybersecurity section opens with a qualifier that changes how its numbers read. "When accessed through the Fairwind Program," according to Google, Carbon "leads on key benchmarks across finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities." The leading scores belong to "Carbon for cyber," the version given to vetted defenders, not necessarily to the model sold through the API.

When Argon launched, Google said Fairwind partners could use it "without cyber guardrails," and that it was expanding the programme to hundreds of additional partners. The Carbon post does not repeat the phrase, does not say whether Carbon for cyber carries guardrails, and does not say how many organisations are now enrolled.

Google lists three early results:

  • Project Fortify, Google's internal system for scanning, triage and remediation: Carbon "hit 100% recall on scanning and 95% precision on triage, on our internal benchmark," and showed "an over 25% relative improvement on Argon's already strong recall rate" in open-source repository testing.
  • Cloud Vulnerability Research: Google's CVR team used Carbon to show that two bugs in a popular open-source project were exploitable vulnerabilities, "comparable in severity to the recent PHP2Shell exploit." The maintainers reclassified the bug as critical and issued a patch; Google says details will follow once disclosure is complete.
  • Wiz Scan for Good: scanning two major SaaS providers, the model "uncovered 0-day vulnerabilities in under two hours," according to the post, which says the fixes helped mitigate potential impact on "tens of thousands of companies."

The PHP reference ties Carbon to a race that has been running in public. On October 21, the NSA and CERT-EU warned about two PHP flaws rated 9.2 that let attackers run code, and OpenAI said the same day that its GPT-5.6 Sol model helped find one of them. Earlier in the year, Mozilla credited Anthropic's restricted Claude Mythos Preview with 271 Firefox security fixes. Google, OpenAI and Anthropic have now each tied a model to the discovery of serious software flaws.

Two security executives are quoted. Chris Corde, chief product officer at SentinelOne, said: "In a controlled evaluation of kernel driver vulnerability research, every model we tested found the same core memory-corruption bugs. Carbon for cyber went further, it understood the driver's constraints, separated the findings that hold up from those that don't, and uncovered four additional verified vulnerabilities." John White, director of the Advanced Security Initiatives Group at Cisco, said deep reasoning models such as Carbon for cyber let defenders judge whether vulnerabilities in complex codebases are genuinely exploitable, and that early testing showed a level of precision that could let engineers "move past the noise."

For advertisers and publishers, the relevance is indirect but real. Much of the web, ad tech included, runs on open-source software of the kind these models scan, and a flaw that a defender finds in two hours might be found as quickly by an attacker with similar tools. Google's own safety framework, revised to version 3.1 in April, recommends higher security levels for models that reach a critical capability level for cyber uplift. The post does not say whether Carbon reached one.

Science, and a programme for 10,000 researchers

The science section reads as a list of research results. According to Google, drug toxicity company Axiom Bio found Carbon to be the leading model for translational science and is moving production analyses and agents to both Carbon and Argon. Google researchers built an agentic formal verification framework called AlphaProof Code and used Carbon, alongside other models, to generate "millions of lines of code in Lean" to prove security properties of the code underlying Signal's cryptographic protocol, which Google says took weeks rather than the years such work historically required.

In enzyme assay development, Carbon worked from lab notebooks and raw data to diagnose experimental artefacts, including incomplete automated mixing in deep wells. In protein structure prediction, across seven low-homolog protein families from the CASP17 blind test set, Carbon "increased the number of homologs retrieved by 40-fold on average," according to Google, which adds that "broader benchmarking is still underway."

Four weeks earlier, alongside Argon, Google had opened applications for 10,000 scientists to use Argon in Antigravity free of charge for six months. The November 10 post says "thousands" have been onboarded and that current and future participants will get Carbon too.

The safety figures

Because Carbon has stronger cyber capabilities, Google says it built on Argon's safety stack to make the model more robust against misuse and prompt injection. The post gives these numbers, all from Google or testers it engaged:

  • In simulated cyberattack environments seeded with ExploitBench and CyScenarioBench, Carbon blocked 99% of attack trajectories from an adaptive red-teaming agent. 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.
  • External testing by Trajectory Labs found "zero new universal jailbreaks across 132 hours and over 13,000 requests."
  • On a 375-task agentic honesty benchmark, Carbon passed 92% of tasks, against 70% for Argon. It declined to fill in missing experimental data 98% of the time and flagged methodological flaws 68% of the time.
  • On the Gray Swan indirect prompt injection benchmark, Carbon recorded a 0% attack success rate, a result Google first shared two weeks earlier and calls a first. Argon's figure on the same benchmark was 0.7% across 15 adaptive attempts.

The honesty numbers deserve a second look from anyone using a model to analyse campaign data. Declining to invent missing data 98% of the time is a strong figure; flagging methodological flaws 68% of the time means roughly one flawed method in three went unflagged on Google's own test.

Google also qualifies its prompt injection result. "At the same time, we know that prompt injection robustness remains an area for continued improvement," the post says, adding that its automated red teaming "has shown that models could be further improved against sophisticated attacks in some distributions." For a model that will read web pages in AI Mode and act inside connected tools, that caveat matters as much as the 0%.

Google points to a separate Gemini 4 Carbon Safety and Security report and a model card for detail.

Why this matters for marketers

Three parts of the release reach marketing work directly.

Search. Google's strongest model now writes AI Mode answers for Ultra subscribers. The audience is small and expensive, but it is the same channel into which Google is pushing ads across more markets, and the post is silent on how Carbon handles citations, links or ad placements.

Cost. For agencies and martech vendors building on the Gemini API, Carbon costs the same per token as Argon. Google's own description, however, says it spends more tokens per task. Google's cost advantage claims compare Carbon with unnamed external models, not with Argon, so the post offers no basis for estimating what moving a reporting or analysis workflow from Argon to Carbon would add to a monthly bill.

Access. Every route to Carbon described in the post runs through a payment or a selection process: the paid API tier, Gemini for business, Ultra, a science programme or Fairwind. The pattern set in October, when free users were cut to Flash-Lite and Argon went to paying tiers, holds.

What the announcement does not provide is as significant as what it does: the identity of the models Carbon beat, Carbon's token consumption relative to Argon, the length of the introductory price, usage limits, AI Mode markets and languages, and any statement on ads.

Timeline

Summary

Who: Google and Google DeepMind, in a post by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google. Partners quoted or cited include Hebbia, Balyasny Asset Management, OpenCode, Cartwheel, SentinelOne, Cisco, Wiz, Axiom Bio, Artificial Analysis, Gray Swan and Trajectory Labs. The release affects developers and enterprises on the paid Gemini API, Google AI Ultra subscribers, and advertisers and publishers whose audiences use AI Mode.

What: Gemini 4 Carbon, the second Gemini 4 model, became available through the Gemini API and Gemini for business at the same introductory price as Gemini 4 Argon, $2 per million input tokens and $10 per million output tokens, and began rolling out to Ultra subscribers in AI Mode, the Gemini app and Google Antigravity. Google says Carbon uses more tokens than Argon, previews a multi-agent /swarm feature in Antigravity, and reports benchmark, cybersecurity and safety figures that are its own or its partners'.

When: November 10, 2026, thirteen days after Google first released Carbon to cybersecurity defenders on October 28 and four weeks after Argon's release on October 13.

Where: Through the Gemini API in Google AI Studio and Android Studio, Gemini for business, CodeMender, Google Antigravity, the Gemini app and AI Mode in Search. The post names no countries or languages for the consumer rollout.

Why: Google positions Carbon for the hardest coding, research, finance and security tasks, with Argon remaining its default for most work. For marketers, Carbon changes the model behind some AI Mode answers for the highest-paying subscribers, and it offers agencies a more capable API model at an unchanged per-token price but a higher token appetite, while the announcement says nothing about ads, citations or Google's advertising tools.