Quick Answer
Gemini 4 Carbon is Google's strongest AI model, generally available since 10 November 2026. It is built for long, complex work such as multi-step coding, agentic workflows and cybersecurity. It costs $2 per million input tokens and $10 per million output tokens until 30 November 2026, then doubles to $4/$20 from 1 December. It is not free: you need the paid Gemini API or a Google AI Ultra subscription.
Carbon is built for agentic work — see what agentic AI is and how it works to understand the agents it powers.
Introduction
Google has a new top model, and it arrived with a deadline. Gemini 4 Carbon went from a limited preview for cybersecurity defenders to general availability on 10 November 2026, and its introductory price runs out on 30 November. After that, the same tokens cost exactly twice as much.
If you build with AI, write code, or use the Gemini app, this matters to you — but probably not in the way the headlines suggest. Google itself says its other model, Gemini 4 Argon, is still the right choice for most work. Carbon is for the hardest, longest jobs.
This guide covers what Gemini 4 Carbon is, what it costs, how it performs, and exactly how to access it — including the routes that do not work.
What Is Gemini 4 Carbon?
Gemini 4 Carbon is a frontier model in Google's Gemini 4 generation. Google calls it its strongest model, designed for work that tests the limits of a model's intelligence: complex, multi-step coding and debugging, open-ended scientific research, and cybersecurity defence.
It is not a new product family. It is the top tier of the Gemini 4 line, above Gemini 4 Argon. Think of Argon as the everyday workhorse and Carbon as the specialist you call in for the biggest jobs.
Carbon's model ID is gemini-4-carbon. It accepts text, image, video, audio and PDF as input, and produces text output. Its knowledge cutoff is March 2026, so it does not know about events after that date.
Three things define Carbon:
- Long-horizon work. It is built to keep going on tasks that take many steps, such as migrating code across several packages, tracking down a bug across subsystems, or redesigning an architecture.
- Build-and-verify loops. For web apps and full-stack work, it writes code, runs it, checks the rendered result, and recovers from its own errors. That loop is the heart of vibe coding — building an app by describing it while the AI writes the code.
- Cybersecurity. It can find, exploit and patch vulnerabilities. That capability is why Google first gave it to a small group of cyber defenders before opening it to everyone else.
Carbon also supports function calling, structured outputs, code execution, computer use, file search and Google Search grounding. It has a unified context window of 1,048,576 tokens (about 1 million), shared across input, thinking and output. The default output limit is 65,536 tokens, but long decode can stretch a single response toward the full 1 million.
You can set its thinking depth to three levels — low, medium (the default) or high. Higher thinking means deeper reasoning, but it also means more tokens, and therefore a bigger bill.
When Was Gemini 4 Carbon Released?
| Date | What happened |
|---|---|
| 28 October 2026 | Announced by Google, given first to a subset of cyber defenders through the Fairwind Program |
| 10 November 2026 | Generally available (GA) for paying developers, enterprises and Google AI Ultra subscribers |
| 30 November 2026 | Introductory pricing ends |
| 1 December 2026 | Price doubles to $4 input / $20 output per million tokens |
For comparison, Gemini 4 Argon was announced on 30 September 2026 and became generally available on 13 October 2026.
Gemini 4 Carbon Pricing
This is the part to read carefully, because the price changes in a matter of days.
| Through 30 Nov 2026 | From 1 Dec 2026 | |
|---|---|---|
| Input (per 1M tokens) | $2.00 | $4.00 |
| Output, including thinking (per 1M tokens) | $10.00 | $20.00 |
| Cached input (per 1M tokens) | $0.10 | $0.20 |
| Batch / Flex (per 1M tokens) | $1.00 / $5.00 | $2.00 / $10.00 |
A few pricing details that catch people out:
- Thinking tokens are billed as output. Carbon uses more thinking tokens per turn than Argon, so the same task can cost more on Carbon even though the per-token price is identical.
- There is no free tier. Carbon is not available on the Gemini API free tier. Every token is billed.
- Caching helps. Implicit caching starts at 4,096 prompt tokens and discounts cached input by 95%.
The practical takeaway: if you want to try Carbon cheaply, do it before 30 November, while the introductory price holds. From 1 December, budget double.
What Can Gemini 4 Carbon Do?
Google highlights four strengths:
Agentic coding across multiple packages. Multi-package migrations, cross-subsystem root-cause investigations, and complex architectural redesigns — the kind of work where a model has to hold a whole codebase in mind.
Full-stack build-and-verify loops. It writes code, runs it, checks the rendered UI, and fixes its own errors, producing complete working applications rather than just snippets.
3D and UI design. Stronger spatial reasoning carries over to game development and creative coding, including 3D scenes and polished interfaces.
Cybersecurity. Finding, exploiting and patching vulnerabilities. Google reports that Carbon's safeguards blocked 99% of attack trajectories from an adaptive red-teaming agent. The dedicated cyber-tuned version remains gated behind the Fairwind Program for defenders.
Google is also previewing /swarm in Antigravity, a feature that turns one prompt into a coordinated team of agents working in parallel workspaces.
Gemini 4 Carbon Benchmarks
Benchmarks are the easiest place to be misled, so here is who produced each number.
Google's own results (from its launch evaluation):
| Benchmark | Gemini 4 Carbon |
|---|---|
| Artificial Analysis Intelligence Index v5.0 | 59 — ranked #1 |
| CursorBench 4.0 (agentic coding) | 60.4% — #1 |
| SEC-bench Pro (finding vulnerabilities) | 88.2% |
| ExploitGym (exploitation) | 54.8% |
| CWE-bench (patching) | 71.0% |
| WebDev Arena | #4, 1768 Elo |
Google also says Carbon achieves more than double the success rate of Argon at generating secure code, and a 0.0% attack success rate on prompt-injection testing, compared with 0.7% for Argon. These are Google's own results, not independent tests.
Independent results (Vals AI, a third-party evaluator):
| Benchmark | Score | Rank |
|---|---|---|
| Vals Index | 70.45% | #2 of 32 |
| Finance Index | #1 | #1 |
| Legal Research Bench | #1 | #1 |
| Tax Agent Bench | #1 | #1 |
| Terminal-Bench 4.0 | 63.13% | #4 of 59 |
One useful comparison from the launch data: on CursorBench 4.0, Carbon scored 60.4% at $9.22 per task, while Claude Fable 5.2 Max scored 60.0% at $32.79 per task. In other words, Carbon matched the top competitor at roughly a third of the per-task cost — at introductory prices.
The honest read: Carbon is genuinely strong on coding, agentic work and security, and independent testing backs up much of Google's claim. But benchmarks are snapshots. Test it on your own work before committing to it at the December price.
Gemini 4 Carbon vs Gemini 4 Argon
This is the comparison most people actually need.
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Position | Strongest model, top tier | Everyday flagship |
| Best for | Hardest, longest jobs; multi-step coding; security | Most coding and knowledge work |
| Price (through 30 Nov) | $2 / $10 per 1M tokens | $2 / $10 per 1M tokens |
| Price (from 1 Dec) | $4 / $20 per 1M tokens | $4 / $20 per 1M tokens |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Thinking tokens | Uses more per turn | Uses fewer per turn |
| In the Gemini app | "Most complex work" | "Code & analysis" |
| Available on | Paid API, Google AI Ultra | Paid API, Google AI Pro and up |
Google's own guidance is clear: 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.
The practical rule: use Argon by default. Switch to Carbon only when a task is long, multi-step, or keeps failing on Argon. Because Carbon thinks more per turn, an identical per-token price can still mean a larger bill.
How to Access Gemini 4 Carbon
There are four routes, and they are not equal.
1. Developers: the paid Gemini API
This is the most direct route. If you have a Gemini API key with billing turned on, change the model string to gemini-4-carbon and send a request. Google says nothing else in the request has to change if you are coming from Argon.
- Get a key and enable billing in Google AI Studio.
- Set the thinking level with
thinking_level:low,medium(default) orhigh. - For long jobs, allow at least 32,768 output tokens and a 15-minute timeout, since Carbon's reasoning takes time.
A paid-tier API key needs no subscription at all.
2. Enterprises: Gemini for business
Available through the Gemini for business API on Google Cloud.
3. Consumers: Google AI Ultra in the Gemini app
Carbon is rolling out to Google AI Ultra subscribers in the Gemini app, in AI Mode in Search, and in Antigravity. In the model picker, 4 Argon is labelled "Code & analysis" and 4 Carbon is labelled "Most complex work". A rollout can take a few days to reach every account, so if you do not see it yet, check again.
4. Cyber defenders: the Fairwind Program
The cyber-tuned version is offered to security teams through Google's Fairwind Program on a rolling basis. This is a separate route with its own application — it is not a prerequisite for ordinary API use.
Routes that do NOT work
- The Gemini API free tier. Carbon is not on it.
- Google AI Pro. A Google AI Pro subscription does not include Carbon in the app. Only Ultra does. (If you have Google AI Pro and want to try Carbon, the paid API is the cheaper route — you pay per token, with no subscription.)
Is Gemini 4 Carbon Free?
No. There is no free way to use Gemini 4 Carbon. It is available only on the paid Gemini API tier (pay per token) or through a Google AI Ultra subscription. The free tier of the API and the Google AI Pro plan do not include it.
If you want a free Google model, Argon is available on the Google AI Pro plan, and the Gemini API free tier covers other models. Carbon is the paid top tier.
Who Should Use Gemini 4 Carbon?
Use Carbon if you:
- Work on large, multi-step coding projects or migrations
- Build agentic workflows or multi-agent systems
- Do security work, such as finding or patching vulnerabilities
- Need the deepest reasoning on a hard problem, and the cost is worth it
Stick with Argon if you:
- Do everyday coding, writing, analysis or chat
- Want the lowest bill — Argon uses fewer thinking tokens per task
- Are just starting with the Gemini API
Skip both and use a free model if you:
- Only need short answers, summaries or simple code
- Are learning and do not want a bill at all
Common Mistakes to Avoid
- Forgetting the 1 December price rise. The introductory $2/$10 rate ends 30 November. Budget for $4/$20 after that.
- Ignoring thinking tokens. They are billed as output, and Carbon uses more of them than Argon. A long reasoning task can cost more than the per-token price suggests.
- Trusting the knowledge cutoff. Carbon's knowledge ends in March 2026. For anything after that, give it sources or use Search grounding.
- Assuming your plan includes it. Google AI Pro does not include Carbon. Check your plan before you look for it in the app.
- Using Carbon for simple tasks. It is slower and more expensive than Argon on easy work. Match the model to the job.
Key Takeaways
- Gemini 4 Carbon is Google's strongest AI model, generally available since 10 November 2026.
- It is built for long, complex, multi-step work — coding, agentic workflows and cybersecurity.
- Pricing is $2/$10 per million tokens until 30 November 2026, then doubles to $4/$20 from 1 December.
- There is no free tier. Access is via the paid Gemini API or a Google AI Ultra subscription. Google AI Pro does not include it.
- Argon remains Google's recommended model for most work. Use Carbon only for the hardest jobs.
- Its knowledge cutoff is March 2026, and thinking tokens are billed as output.
FAQs
What is Gemini 4 Carbon? Gemini 4 Carbon is Google's strongest AI model, released on 10 November 2026. It is designed for long, complex work such as multi-step coding, agentic workflows and cybersecurity, and it is the top tier of the Gemini 4 model family.
Is Gemini 4 Carbon free? No. Gemini 4 Carbon is not available on the free tier. You need the paid Gemini API tier (pay per token) or a Google AI Ultra subscription. Google AI Pro does not include it.
How much does Gemini 4 Carbon cost? Until 30 November 2026, it costs $2 per million input tokens and $10 per million output tokens. From 1 December 2026, the price doubles to $4 input and $20 output per million tokens.
How do I access Gemini 4 Carbon?
Developers can call gemini-4-carbon through the paid Gemini API in Google AI Studio. Google AI Ultra subscribers can select it in the Gemini app, in AI Mode in Search, and in Antigravity. Enterprises can use it through Gemini for business.
What is the difference between Gemini 4 Carbon and Gemini 4 Argon? Both cost the same per token, but Carbon is built for the hardest, longest tasks and uses more thinking tokens per turn. Google recommends Argon for most coding and knowledge work, and Carbon only for tasks that test the limits of a model's intelligence.
What is Gemini 4 Carbon's context window? It has a unified context window of 1,048,576 tokens (about 1 million), shared across input, thinking and output. The default output limit is 65,536 tokens, extendable toward 1 million with long decode.
Does Gemini 4 Carbon work in the Gemini app? Yes, for Google AI Ultra subscribers. In the model picker, 4 Carbon is labelled "Most complex work". It is rolling out, so it may take a few days to appear on every account.
Is Gemini 4 Carbon better than Claude and GPT models? Google reports it ranks #1 on the Artificial Analysis Intelligence Index (59) and on CursorBench 4.0 (60.4%). Independent testing by Vals AI ranks it #2 of 32 models overall, and #1 on finance, legal research and tax benchmarks. Benchmarks are snapshots, so test it on your own work before deciding.
Conclusion
Gemini 4 Carbon is a real step up for the hardest AI work — long coding projects, agents and security — and its introductory price makes November the cheapest month to try it. But it is not a model for everything. For most daily work, Argon is faster, cheaper and recommended by Google itself.
If you want to try Carbon, do it before 30 November while the $2/$10 rate holds, and start with a small task so you can see how many thinking tokens it actually uses.
Want to get more out of AI tools? See our guides to GPT-6 in ChatGPT: What's New, What's Free & How to Access It, Claude Haiku 5.5: Pricing, Benchmarks & How to Use It, and 100 AI Prompts for YouTube Creators.
Sources: Google Gemini Developer API pricing page (ai.google.dev/gemini-api/docs/pricing); Google's Gemini 4 Carbon announcement, 10 November 2026; 9to5Google, "Gemini 4 Carbon rolling out to Gemini app for AI Ultra", 10 November 2026. Benchmarks are as reported by Google and by Vals AI. Prices and availability can change — confirm on the official page before you spend.

