Skip to content
ISSUE 001·LIVE·20:20 IL
← Journal/2026-11-12·9 min·google gemini

Gemini 4 Carbon GA: what paying developers get today, and the November 30 price jump

Google made Gemini 4 Carbon generally available on November 10: paid API access in AI Studio and Gemini for business, a /swarm multi-agent preview in Antigravity, Fairwind cybersecurity access, and introductory pricing that doubles after November 30, 2026. What changed since the October announcement, what Google claims, and what to verify yourself.

By Harel Asaf·AI Builder·Tel Aviv

Google made Gemini 4 Carbon GA on November 10, two weeks after announcing it on October 28. The Gemini 4 Carbon GA release changes three practical things: any paying developer or enterprise can now use it through the API in Google AI Studio and Gemini for business, Google AI Ultra subscribers get it in Antigravity, the Gemini app and AI Mode in Search, and the introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens doubles to 4 and 20 dollars after November 30, 2026. Google is also previewing /swarm, a multi-agent mode in Antigravity built around Carbon. If you plan to evaluate it, the next two and a half weeks are the cheap window.

What the Gemini 4 Carbon GA release actually changes

At the October announcement, Carbon was limited to cybersecurity defenders in Google's Fairwind Program. Per Google's availability post by Koray Kavukcuoglu, SVP at Google DeepMind, the rollout now covers:

  • Developers: the Gemini API paid tier in AI Studio and Android Studio. Google AI Ultra subscribers can also use it inside AI Studio and Antigravity.
  • Enterprises: Gemini for business via the API or the model dropdown, Google Antigravity, and CodeMender.
  • Consumers: the Gemini app and AI Mode in Search, for Google AI Ultra subscribers.
  • Scientists: the free six-month Antigravity access program that opened to 10,000 researchers four weeks ago now extends to Carbon, with rolling weekly onboarding.
  • Cybersecurity defenders: the Fairwind Program continues on rolling application.

Positioning per Google: Argon remains the default for the majority of software engineering and knowledge work, while Carbon is built for tasks at the limit of model intelligence, like complex multi-step coding and debugging or open-ended research, and deliberately uses more tokens to explore, reason, and run extended verification loops. Read that last part together with the pricing.

Pricing: the November 30 cliff

Carbon launches at the same introductory price as Gemini 4 Argon: 2 dollars per million input tokens, 10 dollars per million output tokens. A footnote in Google's post says that after the introductory period expires on November 30, 2026, the price becomes 4 dollars per million input tokens and 20 dollars per million output tokens, for the public-access version of the model.

For agent workloads the doubling is amplified by design: a model positioned to spend more tokens per task, billed at twice the output rate, can cost well over twice as much per completed task as it does today. If you benchmark it this month, record token consumption per task, not just the headline rates, or your December bill will not match your November test.

Benchmarks: vendor-reported, with one new benchmark

Google says Carbon ranks first 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. Because existing benchmarks do not capture agentic work over dense multimodal documents, Google partnered with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark, where Carbon ranks second at more than 6x lower average cost per task than the top performer. In pre-release evaluations Google also cites matching the best tested model on Hebbia's E2E Financial Agentic evaluation and ranking first out of 23 frontier configurations on Balyasny Asset Management's General Financial Intelligence benchmark.

These are vendor and partner-reported results, and MKA is a brand-new benchmark co-created with the vendor's partner. Directional, not settled.

/swarm: multi-agent mode in Antigravity

The most interesting developer feature in the release is /swarm, now in preview in Google Antigravity: one prompt becomes a coordinated team of agents you track in real time, running in parallel isolated workspaces with adversarial verifiers that test every change. Google's suggested pattern is assigning Carbon to the deep-reasoning roles (architectural design, coordination, auditing) and pairing it with faster, cheaper models for execution.

Google's demos are vendor demos, but they show the intended shape: a team of Carbon agents turned seven public City Council planning PDFs with 206 extracted engineering drawing views into a 906K-triangle interactive Three.js reconstruction across 10 overnight loops, and built an open-world 3D browser game coordinating subagents for Blender assets, WebGL shaders, and generated music. /swarm is a preview gated to Google AI Ultra subscribers.

Cybersecurity and the Fairwind Program

Accessed through Fairwind, Carbon leads Google's cited benchmarks for finding (SEC-bench Pro), exploiting (ExploitGym), and patching (CWE-bench v1) vulnerabilities. Google reports its internal Project Fortify hit 100% recall on scanning and 95% precision on triage, with over 25% relative recall improvement over Argon on open-source repository testing. Its Cloud Vulnerability Research team used Carbon to prove two bugs in a popular open-source project were exploitable vulnerabilities, which maintainers re-classified as critical. Wiz reports that in its Scan for Good program, Carbon surfaced 0-day vulnerabilities at two major SaaS providers in under two hours each. All of these are Google or partner accounts; the Project Fortify figures are internal benchmarks.

Safety numbers worth reading

The safety section is unusually specific for a launch post. Google reports: a 0% prompt injection attack success rate on Gray Swan's IPI benchmark, a first for any model, while cautioning that its own red teaming still finds room for improvement; 99% of attack trajectories blocked from an adaptive red-teaming agent, with the per-attempt breach rate down to 0.8% from 3.1% on Argon; automated cyberattack query success down to 2.17% from 9.86% in Gray Swan's testing; zero new universal jailbreaks found by Trajectory Labs across 132 hours and over 13,000 requests; and 92% pass rate on Google's 375-task agentic honesty benchmark versus 70% for Argon, including declining to fabricate missing experimental data 98% of the time.

What I would check

My untested suggestions before adopting it:

  • Benchmark before November 30, but model your costs at the post-introductory 4 and 20 dollar rates; that is the price you will actually pay at scale.
  • Run Carbon against Argon on your own tasks. Google itself says Argon is the right default for most engineering work, so Carbon has to win by a margin on your hardest tasks to justify doubling cost.
  • Treat /swarm as a preview: test it on a scoped project with verifiable output before trusting overnight autonomous loops.
  • Reproduce the benchmark claims that matter to you. CursorBench, WANDR and AutomationBench leads are Google's reporting, and MKA is new.
  • If you do security work, apply to Fairwind now; access is rolling and gated, not self-serve.

What I would not conclude

Do not conclude Carbon is the best model for your stack; the ranking evidence is vendor-run. Do not treat Argon as deprecated; Google explicitly positions it as the default. Do not budget at introductory pricing past November 30. Do not assume GA means everywhere: consumer surfaces are gated to AI Ultra subscribers, and /swarm is a preview. And do not read 0% on one prompt injection benchmark as immunity, when Google itself says robustness still needs work.

Frequently asked questions

What changed with the Gemini 4 Carbon GA release?

On November 10, Google opened Gemini 4 Carbon to all paying developers and enterprises through the API in AI Studio and Gemini for business, and began rolling it out to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search. At the October 28 announcement it was restricted to Fairwind Program cybersecurity defenders.

How much does Gemini 4 Carbon cost after November 30, 2026?

Per Google's footnote, after the introductory period expires on November 30, 2026, Carbon costs 4 dollars per million input tokens and 20 dollars per million output tokens, double the introductory 2 and 10 dollar rates it shares with Gemini 4 Argon.

Where can developers use Gemini 4 Carbon today?

Through the Gemini API paid tier in Google AI Studio and Android Studio. Google AI Ultra subscribers additionally get it in AI Studio and Google Antigravity, and enterprises can reach it via Gemini for business, Antigravity, or CodeMender.

What is /swarm in Google Antigravity?

A previewed multi-agent mode: a single prompt spins up a coordinated team of agents in parallel isolated workspaces, with adversarial verifiers testing changes and real-time tracking. Google's suggested setup uses Carbon for deep-reasoning roles like architecture and auditing, paired with lighter models for execution. It is a preview for AI Ultra subscribers.

Should I switch from Gemini 4 Argon to Carbon?

Google positions Argon as the right choice for most software engineering and knowledge work, and Carbon for the hardest multi-step problems where its deeper reasoning and extended verification loops justify using more tokens. After November 30, Carbon also costs twice as much per token, so the switch only makes sense where it measurably outperforms Argon on your tasks.

How do I get Gemini 4 Carbon for cybersecurity work?

Apply to Google's Fairwind Program, which grants rolling access to Carbon for cyber. Google says this access path leads its cited finding, exploiting, and patching benchmarks (SEC-bench Pro, ExploitGym, CWE-bench v1).

Are Gemini 4 Carbon's benchmark and safety claims verified?

They are vendor and partner-reported. Google cites first place on CursorBench 4.0, WANDR, and AutomationBench, a 0% prompt injection success rate on Gray Swan's IPI benchmark, and 92% on its own agentic honesty benchmark. Independent evaluations have not appeared yet, and Google itself notes prompt injection robustness remains an area for improvement.

Sources

  • Gemini 4 Carbon is now available | Google Blog - November 10, 2026, read November 12, 2026. Availability, positioning versus Argon, benchmark claims, /swarm preview, Fairwind Program, safety and security figures, science program expansion, and the November 30 pricing footnote.
  • Gemini 4 Carbon: New model for developer agents | Google Blog - November 10, 2026, read November 12, 2026. Developer availability through AI Studio and Android Studio, and Antigravity access for AI Ultra subscribers.
  • TLDR AI newsletter, November 11, 2026 - the tip that surfaced the GA release.

Not read: the Gemini 4 Carbon model card, the full Safety and Security report, the Gemini 4 developer guide, and the Antigravity blog's /swarm details. All benchmark, customer, and safety figures are Google's or its partners'; I did not test the model, and /swarm is a preview limited to AI Ultra subscribers.

Related reading

  • Gemini 4 Hydrogen: Google's fast execution model undercuts Argon on price and beats it on intelligence

    Google launched Gemini 4 Hydrogen on November 12: a fast everyday coding and agentic model that scores 51 on the Artificial Analysis Intelligence Index at its default thinking level, above Gemini 4 Argon's 50, at 1.50 dollars per million input tokens and 7.50 per million output tokens. What the three-model Gemini 4 ladder means, how Orchestrator splits work in Antigravity, and what to verify.

  • Qwen4-Max: Alibaba's 2.4T flagship API model claims 62% lower cost per task than its predecessor

    Alibaba's Qwen team released Qwen4-Max, the first model in the Qwen4 family: a 2.4-trillion-parameter sparse MoE with 71B active parameters, a 1M-token context window, thinking modes with tunable effort, and both OpenAI-compatible and Anthropic-compatible endpoints. Vendor-reported numbers show up to 49% fewer tokens on agentic coding tasks and a 62% drop in average cost per task versus Qwen3.8-Max. Input starts at 1.5 dollars per million tokens.

  • Inkling 2 Small: Thinking Machines ships a 594B open-weights MoE that runs on 17B active parameters

    Thinking Machines Lab released Inkling 2 Small with open weights under Apache 2.0: a natively multimodal 594B mixture-of-experts model with 17B active parameters, 1M tokens of context, and agentic-coding benchmarks that beat most open flash-class rivals. Self-hosting needs a GPU cluster; Ollama Cloud offers it at 0.30 dollars per million input tokens.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS