Google has opened the vault. Gemini 4 Carbon, which the company announced in late October but gave only to a small circle of cybersecurity defenders, became available on November 10 to paying developers, enterprises and Gemini Ultra subscribers. Google calls it “our strongest model,” built for “long-horizon agentic workflows across complex knowledge work, software engineering, and scientific reasoning”, and it is launching at a price that undercuts OpenAI’s and Anthropic’s top models by a wide margin.
Quick Take
- What changed: Carbon moved from a gated cyber-defender preview (via Google’s Fairwind Program) to general availability on the paid Gemini API, Gemini for business and for Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Google Antigravity.
- Price: $2 per million input tokens and $10 per million output tokens until November 30, 2026; then $4/$20. Google said at announcement that cached input is 95% off the input price.
- 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 compared to the next best external model.” Cursor reports 60.4% on CursorBench at high effort.
- New tool: A preview of
/swarmin Google Antigravity turns one prompt into a coordinated team of agents, with Carbon in the planner and auditor roles. - Not for everyone yet: No access for Gemini free or lower paid consumer tiers; cyber-specific capabilities stay behind the Fairwind Program.
What happened
In a blog post by Koray Kavukcuoglu, SVP at Google DeepMind and Google’s Chief AI Architect, Google wrote: “Today, we’re making Gemini 4 Carbon available to developers, enterprises, scientists, and consumers.” Sundar Pichai said it is “built for sustained, long-horizon tasks, reasoning deeply through complex software engineering, cybersecurity, and scientific discovery.” Google’s Logan Kilpatrick added that “this model makes significant progress on lots of the feedback from Argon.”
Google positions the two Gemini 4 models as a pair: “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”: and it uses “more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.” Cursor added Carbon the same day and said that “at high effort, it’s the top model on CursorBench at 60.4%.” Cursor’s chart puts that run at roughly $9.22 per task, with Claude Fable 5.2 just below on score.
Background: from cyber vault to general release
When Google first introduced Carbon, it held the model back: “Starting this week, we will make Carbon available to a subset of cyber defenders,” Google wrote, while it gathered feedback “before making Carbon available to developers, businesses, and consumers.” Google also said it was taking part in “the U.S. government’s voluntary process for pre-release model testing.” i10x covered that gated launch in Google Releases Gemini 4 Carbon for Cybersecurity Tasks. The general release comes one day after Anthropic shipped Claude Fable 5.2, and days after leaks about further Google models (Gemini Krypton and Hyperion).
How it works and where to get it
The model card lists the distribution channels: the Gemini app, Gemini API, Google AI Studio, Gemini for business, Google Antigravity, AI Mode in Search, and the Fairwind Program “for verified cybersecurity defenders.” Google’s blog details who gets what:
- Developers: paid tier of the Gemini API in AI Studio and Android Studio; Ultra subscribers can also use it in Antigravity and AI Studio.
- Enterprises: Gemini for business via API or model drop-down, Google Antigravity, or CodeMender.
- Consumers: Gemini app and AI Mode in Search, for Google AI Ultra subscribers only.
- Scientists: can apply for six months of free access to Argon and Carbon in Antigravity’s highest usage tier.
- Cyber defenders: Carbon “for cyber” via the Fairwind Program.
The headline workflow feature is /swarm in Google Antigravity, which Google says turns “a single prompt into a coordinated team of agents you can track in real time.” Developers can give Carbon “roles that demand the deepest reasoning, like architectural design, coordination, and auditing,” while lighter models work in parallel and “adversarial verifiers” test every change. Pichai showed it off with a playable game: “Builder and critic subagents ran iterative loops across Blender and Three.js – checking their own renders to build a playable 3D game of a pelican biking around our Mountain View campus.”
| Benchmark (model card, Oct. 2026) | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Artificial Analysis Index v5.0 | 59 | 50 | 57 | 58 |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% |
| OSWorld-2.1 (offline subset) | 81.5% | 74.7% | 80.9% | 87.2% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% |
| GraphWalks 256k-1M (F1) | 96.2% | 84.2% | 71.8% | 66.8% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% |
| ExploitGym (via Fairwind) | 54.8% | 42.0% | 42.4% | 34.5% |
The table shows Google’s own model-card figures as of October 2026; cyber scores reflect access through the Fairwind Program. Carbon does not lead everywhere: GPT-6 Astra is ahead on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1.
Safety
Google says it hardened safeguards because Carbon “introduces stronger cyber capabilities.” In simulated cyberattack environments, Carbon “blocked 99% of attack trajectories” from an adaptive red-teaming agent and cut the per-attempt breach rate to 0.8%, down from 3.1% on Argon. External tester 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 versus 70% for Argon. Google also notes that “prompt injection robustness remains an area for continued improvement.”
Real-world claims
Google lists several early uses: in its Project Fortify security system, Carbon “hit 100% recall on scanning and 95% precision on triage” on an internal benchmark; Wiz used it to find 0-day vulnerabilities in two major SaaS providers “in under two hours.” Carbon also powered DeepMind’s AlphaProof Code effort to formally verify Signal’s protocol code, covered in DeepMind AlphaProof Code verifies Signal protocol.
Pros and cons
For users
- Pro: Ultra subscribers get Google’s strongest model directly in the Gemini app and AI Mode in Search.
- Con: Everyone else waits; The Verge notes “it is not clear when — or if — Carbon will be available to lower tier subscribers or free users.”
For developers
- Pro: Frontier-class results at $2/$10 per million tokens through November, plus a /swarm preview for multi-agent work.
- Con: Prices double on December 1, and Carbon deliberately spends more tokens than Argon, so cost per task can rise even if price per token looks low.
For companies
- Pro: Strong scores on AutomationBench and finance evaluations (Hebbia, Balyasny), plus access via Gemini for business and CodeMender.
- Con: Most headline numbers come from Google or partners; full cyber capabilities require separate Fairwind approval.
What X is saying
Gemini 4 Carbon is now available. It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning. If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
— Google (@Google) November 10, 2026 · View on X
Gemini 4 Carbon is now available in Cursor! At high effort, it’s the top model on CursorBench at 60.4%
— Cursor (@cursor_ai) November 10, 2026 · View on X
Our latest frontier model Gemini 4 Carbon is rolling out today! Amazing progress from Gemini 3 → Gemini 4 in less than a year… Carbon reverse engineers my original Theme Park binaries, adds a new DeepMind themed ride, and ports it to the web in one shot – kind of mindblowing to see!
— Demis Hassabis (@demishassabis) November 10, 2026 · View on X
The pelican demo became the launch’s meme. @agenticjay, who built the demo, wrote: “I took @simonw’s benchmark ‘pelican riding a bicycle’ way too far using Gemini 4 Carbon and a /swarm in @antigravity to build a playable game.” Not everyone was charmed: @DirtyTesLa replied to Pichai’s demo: “So thankful RAM is $500 per GB for this.”
What Grok says
Asked how Carbon compares with Anthropic’s newest model, Grok answered: “On the Artificial Analysis Intelligence Index, Claude Fable 5.2 (max) scores 65 while Gemini 4 Carbon (high) scores 60. Fable 5.2 also leads most shared agentic coding, science, and computer-use benchmarks. Carbon is cheaper and stronger on some automation and cyber tasks.” In another reply, Grok summed up: “Carbon far cheaper ($2-4/$10-20 vs $8/$40).” Note that Google’s model card lists Carbon at 59 on the index; Grok cites a later score of 60.
Critical analysis
Carbon’s real pitch is price-performance, not an outright crown. On Google’s own table it trails GPT-6 Astra on FrontierSWE 2 and Opus 5.5 on OSWorld-2.1, and Grok’s reading of independent scores puts Fable 5.2 ahead overall. What Google can claim is a frontier-class model at roughly a fifth of Astra’s token price. The Verge calculates “around a fifth the cost of OpenAI’s GPT-6 Astra and a quarter the cost of Anthropic’s Fable 5.2.”
Two caveats matter. First, the introductory price ends on November 30; at $4/$20 the gap narrows. Second, Google itself says Carbon uses more tokens than Argon to “explore thoroughly,” so token prices are a poor proxy for real bills. The Verge makes the same point: lower per-token prices do not “necessarily translate to lower costs.” Cursor’s chart, which shows cost per task, is the more useful number, and it puts high-effort Carbon above $9 per task.
The staged rollout also sets a precedent. Google released a frontier model to vetted cyber defenders first and kept the strongest cyber capabilities behind an application process even at general release. That is the same pattern Anthropic used with Mythos and OpenAI with its cyber model (OpenAI cyber model).
What’s still open
- Will Carbon reach Google AI Pro or free Gemini users, and when?
- How will real-world cost per task compare after December 1, when prices double?
- Will independent evaluators confirm Google’s AutomationBench and WANDR results?
- When does /swarm leave preview, and what will it cost at scale?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Google’s November 10 blog posts, the Gemini 4 Carbon model card, The Verge’s reporting, and public posts on X through November 11, 2026. i10x did not test Gemini 4 Carbon directly.
i10x Perspective
Within 48 hours, Anthropic and Google both shipped their strongest models, and both told developers the same thing: use the big model only for the hardest steps. Carbon’s /swarm makes that explicit (one expensive planner, many cheaper workers, adversarial checkers in between). For teams, the winning move is not picking a single “best” model but measuring cost per finished task across a mix. Google’s aggressive introductory pricing is an invitation to run that test now, before December 1.






