Gemini 4 Carbon Is Out: Swarm, Pricing & Access
We may earn a commission if you buy through links on this page. Read the affiliate disclosure · How we review tools
Quick answer. Gemini 4 Carbon became generally available on Nov 10, 2026. The paid Gemini API price is $2 per million input tokens and $10 per million output tokens through Nov 30, 2026, then $4 and $20. It is not on the free tier. Google is previewing /swarm in Antigravity. We did not run the model.
Gemini 4 Carbon is out. Google made it generally available on Nov 10, 2026, for paying developers and enterprises, and for Google AI Ultra subscribers in the Gemini app, AI Mode in Search, and Google Antigravity. The API id is gemini-4-carbon. On the paid Gemini Developer API the introductory price is $2 per million input tokens and $10 per million output tokens through Nov 30, 2026, then $4 and $20. There is no free tier.
This page is based on official documentation, benchmarks and public reports. We did not run Gemini 4 Carbon, /swarm, or the benchmarks Google cites.
How this page was checked
We read Google’s Nov 10, 2026 availability post, the developer post from the same day, the Gemini 4 Carbon developer guide published Nov 10, 2026, the Gemini Enterprise Agent Platform model page (last updated Nov 13, 2026), and the Gemini Developer API pricing page. The Verge’s Nov 10, 2026 report is used for what it adds as a public account, and it is labeled as The Verge. Claude Fable 5.2 prices and Anthropic’s benchmark cells are from the Nov 9, 2026 Anthropic post and Anthropic’s pricing page, written up in Claude Fable 5.2. Hydrogen and Qwen dates and prices point at the pages already on this site.
Carbon sits above Gemini 4 Hydrogen in Google’s own lineup. Hydrogen is the fast coding model, with a free tier. Both are on the AI model release timeline and in the best AI models ranking.
Key facts
| Gemini 4 Carbon | |
|---|---|
| Public release | Nov 10, 2026. Google describes a Fairwind cyber release about two weeks earlier and does not print that earlier date. |
| Maker | Google DeepMind. The availability post is bylined Koray Kavukcuoglu. |
| API model id | gemini-4-carbon |
| Cyber model id | gemini-4-carbon-cyber-permissive, allowlisted through the Fairwind Program. Computer use is not supported on that id. |
| Role | Google calls it the strongest Gemini 4 model, for cybersecurity and long-horizon work. Argon stays the default recommendation for most software engineering and knowledge work. |
| Context | A unified 1,048,576-token window that counts input and output together, including thinking and function calls, on the Enterprise model page. The developer guide describes the same window as 1M tokens on the Gemini API. |
| Max output | The pricing page says long-decode generation up to 1M tokens. The developer guide says the Interactions API defaults to 65,536 tokens, and that background=True can go past that, up to 1M minus the input. The Enterprise page says one foreground segment is up to 65,536 tokens when streaming and 32,768 when not streaming. |
| Thinking levels | low, medium, high. Medium is the default on gemini-4-carbon. High is the default on the cyber id. |
| Knowledge cutoff | March 2026, on the Enterprise model page. |
| Paid API price | $2 / $10 per million tokens through Nov 30, 2026, then $4 / $20. Output includes thinking tokens. |
| Cache | Implicit only. $0.10 per million through Nov 30, 2026, then $0.20. The developer guide calls that 95% off cached input. Explicit caching is not supported. |
| Free tier | Not available. |
| Regions on the Enterprise page | global, us, and eu, for verified enterprise customers in select regions. The cyber id lists global and us. |
Agent swarm
/swarm is a preview in Google Antigravity, not a flag we found on the public Gemini API price card. Google’s Nov 10 post says it turns a single prompt into a coordinated team of agents you can track in real time. Carbon is the model Google tells developers to put on the roles that need the deepest reasoning: architectural design, coordination, and auditing. Faster models run beside it in isolated workspaces, with verifiers that test each change.
The developer guide gives the same split. Use gemini-4-carbon as the lead orchestrator and verifier. Hand parallel implementation to gemini-4-argon or gemini-3.8-flash. The guide also says Carbon can coordinate other Google models for images, video, and audio. Those are Google’s workflow notes. The demos in the announcement, including a multi-agent 3D reconstruction and a game built from one prompt, are Google’s examples. We did not open them.
Antigravity access in the Nov 10 post is for Google AI Ultra subscribers. Scientists already in the Argon research program, and future applicants, can be offered Carbon in Antigravity’s highest usage tier at no charge for six months. Google says that program is subject to eligibility, capacity, and selection. It is not the public API.
The Verge’s Nov 10, 2026 story describes the same public rollout and the same swarm preview. It also says Google has not made clear whether Carbon will reach lower-tier or free subscribers. Google’s own post is more definite about who has it now: paying API customers, enterprises, and Google AI Ultra on the consumer surfaces. It does not announce a free-tier date.
Gemini 4 Carbon API pricing
Checked on the Gemini Developer API pricing page on Nov 16, 2026. Carbon and Argon use the same introductory schedule. Hydrogen does not.
| Free tier | Paid, per 1M tokens | |
|---|---|---|
| Input | Not available | $2.00 through Nov 30, 2026. $4.00 from Dec 1, 2026 |
| Output, including thinking tokens | Not available | $10.00 through Nov 30, 2026. $20.00 from Dec 1, 2026 |
| Context caching | Not available | $0.10, then $0.20. Implicit only |
| Grounding with Google Search | Not available | 5,000 requests a month free, shared across Gemini 3 and newer models, then $14 per 1,000, with a daily cap of 1,500 |
| Used to improve Google’s products | Not available | No |
Google says every token is billed, including thinking tokens and the tool instructions added when you use grounding. The Carbon price card does not print a separate batch rate. The paid-tier summary on the same page lists a Batch API at a 50% cost reduction as a paid feature, and the Enterprise model page says batch inference is supported. This article does not turn that summary line into a Carbon batch price.
On the Enterprise model page, implicit caching has a 4,096-token minimum, and durable caching is on by default. thinking_budget and thinking_level="MINIMAL" return an HTTP 400 on that surface. The Gemini API developer guide, written for AI Studio, says thinking_budget is deprecated and returns a warning, and that custom temperature, top_p, and top_k are ignored. The two docs do not describe the sampling error the same way. Use the surface you actually call.
Argon matches the $2 / $10 introductory card and the later $4 / $20 card. Gemini 4 Hydrogen is $1.50 / $7.50 on the paid tier, with a free tier, and that rate is not on this introductory clock. Gemini 3.8 Flash remains $0.75 / $3.75 through Dec 31, 2026 on the same pricing page.
What Google says it is for
Google’s post calls Carbon its strongest model, for cybersecurity and for long-horizon agents across knowledge work, software, and science. Argon remains, in Google’s words, the ideal choice for the majority of software engineering and knowledge work. Carbon is the model that spends more tokens to search harder, check longer, and keep a verification loop running. The developer guide says to pick Carbon for the hardest multi-step engineering, architecture and audit, formal math and science, high-precision finance or legal synthesis, and vulnerability research, and to pick Argon when tighter token use and a faster turn matter more.
The benchmark claims Google prints without a raw score:
- Number 1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than 2 times lower cost per task than the next best external model. The post does not print those scores.
- Number 2 on the Multimodal Knowledge Agent benchmark Google says it launched with Artificial Analysis, at more than 6 times lower average cost per task than the top model. The score is not in the post.
- In pre-release tests, matched the best model Google tested on Hebbia’s E2E Financial Agentic evaluation, and ranked at the top of 23 frontier configurations on Balyasny’s General Financial Intelligence benchmark. No numeric score is printed.
- On an internal 375-task agentic honesty benchmark, passed 92% of tasks, compared with 70% for Argon.
Anthropic’s Nov 9 table, published the day before, does print Carbon cells. They are Anthropic’s measurements, not Google’s. FrontierCode 2.0 is 41.1% at Medium effort, FrontierSWE v2 is 61.2%, AutomationBench is 53.9%, Terminal-Bench-Science 0.1 is 64.8%, OSWorld 2.1 is 81.5%, Chartography is 72.5%, and AA-Briefcase is 61.8% on the rubric and 1611 Elo. The full row, including Fable 5.2, is in Claude Fable 5.2. Google’s “number 1 on AutomationBench” claim and Anthropic’s 53.9% cell are left side by side. The posts do not show they are the same run.
Fairwind is the cyber path. On the Enterprise page, Carbon for cyber is generally available behind an allowlist. Access to earlier Gemini cyber models does not carry over. Google’s availability post says that, through Fairwind, Carbon leads on SEC-bench Pro, ExploitGym, and CWE-bench v1. It does not print those scores in the post we used.
Gemini 4 Carbon vs Claude Fable 5.2 and other rivals
List prices checked Nov 16, 2026. Terminal-Bench versions are not one ladder. Hydrogen’s score is Google’s Terminal-Bench 4.1 figure from the Hydrogen article. OpenAI’s Terra score is Terminal-Bench 4.0.
| Gemini 4 Carbon | Gemini 4 Argon | Gemini 4 Hydrogen | Claude Fable 5.2 | GPT-6.1 Terra | Qwen4-Max | |
|---|---|---|---|---|---|---|
| Released | Nov 10, 2026 | The Nov 10 post says Argon launched recently and does not print the date | Nov 12, 2026 | Nov 9, 2026 | Oct 26, 2026 on OpenAI’s research index | Nov 16, 2026 |
| API id | gemini-4-carbon | gemini-4-argon | gemini-4-hydrogen | claude-fable-5-2 | gpt-6.1-terra | qwen4-max |
| Input / 1M | $2.00, then $4.00 from Dec 1 | Same schedule | $1.50 | $8 | $1 | $1.50 in Singapore |
| Output / 1M | $10, then $20 from Dec 1 | Same schedule | $7.50 | $40 | $5 | $4.50 in Singapore |
| Free API tier | No | No | Yes | The Claude API card we read is paid | not published as a free API tier | A Singapore free quota is described on the Qwen page |
| Context | 1,048,576 combined, Enterprise page | Shared unified 1M window on the Gemini API, per the Carbon developer guide | 1M, one rate for the window | 1M | not published by OpenAI | 1,000,000 |
| Terminal-Bench | not printed in Google’s Nov 10 post | not printed there | 57% on 4.1, against 18.2% for Gemini 3.8 Flash (Google) | not in Anthropic’s Fable table as Terminal-Bench. See the Fable article for Terminal-Bench-Science | 50.3% on 4.0 at maximum reasoning (OpenAI) | not published |
| CursorBench, WANDR, AutomationBench | Google says number 1, and does not print the scores | not printed in the Carbon post | not printed in the Carbon post | Anthropic prints WANDR relaxed at 82.1% for Fable and leaves Carbon blank. AutomationBench in Anthropic’s table is 53.9% for Carbon and 51.4% for Fable | not in those posts | not published |
Qwen4-Max is the flagship Alibaba released on Nov 16, 2026. Its benchmark table does not overlap with Google’s Carbon claims, so this page does not rank them on quality.
👍 Pros
- Introductory paid rate of $2 / $10 through Nov 30, 2026, against $8 / $40 for Claude Fable 5.2
- Documented on the paid Gemini API, Gemini for business, and Google AI Ultra surfaces, with a 1,048,576-token combined window
- Google's /swarm preview is a concrete multi-agent mode, with Carbon named as the coordinator
👎 Cons
- No free tier. The price doubles on Dec 1, 2026, to $4 / $20
- Google's number-1 claims do not include the underlying scores in the post we read
- Consumer access in the Gemini app, Search AI Mode, and Antigravity is the Ultra tier in Google's post
Who should upgrade
Use Carbon when the job matches Google’s description of the limit cases: multi-step debugging, long verification loops, dense multimodal files, formal science, or defensive cyber work you are allowed to run. The introductory rate lasts through Nov 30, 2026. After that the card is $4 / $20, which is still half of Fable 5.2’s $8 / $40 list price per token, and still more than Hydrogen.
Stay on Argon for most software engineering and knowledge work. Google says Carbon spends more tokens than Argon on purpose. The same token price during the introductory window can still produce a larger bill if the run is longer. The developer guide’s routing note is to put Carbon on architecture, coordination, and the final check, and to put Argon or Gemini 3.8 Flash on the parallel slices.
Stay on Gemini 4 Hydrogen for everyday coding agents, the free tier, and the flat $1.50 / $7.50 rate. Hydrogen is not a lesser Carbon. Google’s Hydrogen post describes it as the fast model beside Carbon and Argon.
Move to Claude Fable 5.2 when you need Anthropic’s general-release frontier model, or Mythos 5.2 if the work is inside a verification program. Fable’s list price is higher. Anthropic’s table and Google’s rank claims do not share one scoreboard, so this page does not call a winner.
How to use Gemini 4 Carbon
Google’s getting-started list, from the Nov 10 post and the developer guide:
- Paid Gemini API: model id
gemini-4-carbonin Google AI Studio and Android Studio. The developer guide also documents the Interactions API and managed agents that take this model id. Thinking islow,medium, orhigh, withmediumas the default. - Gemini for business: the API or the model menu. The announcement also names Google Antigravity and CodeMender for enterprises. The post does not describe CodeMender beyond listing it.
- Gemini app and AI Mode in Search: Google AI Ultra subscribers.
- Google Antigravity: Google AI Ultra, including the
/swarmpreview. The desktop app, CLI, and IDE extensions are named in the developer guide. - Scientists: the six-month Antigravity program, on a rolling application, subject to eligibility, capacity, and selection.
- Cybersecurity defenders: apply to the Fairwind Program for
gemini-4-carbon-cyber-permissive. - Files: the Enterprise page says Carbon accepts PDF, plain text, Microsoft Office, and OpenDocument files, plus image, audio, and video, within the limits on that page. The Gemini Live API is listed as not supported.
We are not publishing a sample request. The developer guide’s migration notes, if you are coming from Gemini 3 or earlier, are to replace thinking_budget with thinking_level, stop sending custom sampling parameters, and stop using explicit cached_content.
The Enterprise model page limits standard Carbon to global, us, and eu, for verified enterprise customers in select regions. It does not list a China region. The Gemini API pages we read do not name a mainland China endpoint or an approved third-party API router. This article does not add one.
Related AI tools
Carbon is a model. These are separate products for the work around it.
- Cursor for repository coding in an editor. Antigravity’s swarm is Google’s product, and Cursor’s model menu is its own.
- Fireflies for meeting notes and transcripts.
- ElevenLabs if the job is voice. The Carbon guide mentions Google’s own speech and music models. It does not include this product.
Sources
Read on Nov 16, 2026, unless the page prints its own date.
- Gemini 4 Carbon is now available, Nov 10, 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/
- Gemini 4 Carbon for developers, Nov 10, 2026: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-4-carbon-developers/
- Gemini 4 Carbon developer guide, published Nov 10, 2026: https://aistudio.google.com/learn/gemini-4-carbon-developer-guide
- Gemini Enterprise Agent Platform, Gemini 4 Carbon, updated Nov 13, 2026: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-carbon
- Gemini Developer API pricing: https://aistudio.google.com/docs/pricing
- The Verge, Nov 10, 2026: https://www.theverge.com/ai-artificial-intelligence/1030787/google-gemini-4-carbon-public-rollout
- Claude Fable 5.2 announcement and pricing are linked from Claude Fable 5.2.
- GPT-6.1 Terra: https://openai.com/index/introducing-gpt-6-1-terra/
Pricing and availability last checked Nov 16, 2026. This page contains affiliate links. A commission does not change the model facts.
Frequently asked questions
When was Gemini 4 Carbon released?
Google's public-availability post is dated Nov 10, 2026. The Gemini Enterprise Agent Platform model page lists gemini-4-carbon as GA on the same date. Google says a narrower release to cybersecurity defenders in the Fairwind Program happened about two weeks earlier, and that post does not print the earlier calendar date.
What is Gemini 4 Carbon API pricing?
On the Gemini Developer API pricing page, checked Nov 16, 2026, paid input is $2.00 per million tokens through November 30, 2026 and $4.00 from December 1, 2026. Paid output, including thinking tokens, is $10.00 and then $20.00. Implicit context caching is $0.10 and then $0.20. There is no free tier. Explicit caching is not supported.
What is the Gemini 4 Carbon agent swarm?
Google is previewing /swarm in Google Antigravity. The Nov 10 post says a single prompt becomes a coordinated team of agents that you can track in real time. Google says developers can assign Carbon to architecture, coordination, and auditing, and pair it with faster models in parallel workspaces. The consumer and Antigravity rollout described in that post is for Google AI Ultra subscribers.
How do I use Gemini 4 Carbon?
Google's developer guide says to call the model id gemini-4-carbon on the paid Gemini API in Google AI Studio, in Android Studio, and in the Interactions API. Google AI Ultra subscribers can use it in the Gemini app, AI Mode in Search, and Google Antigravity. Enterprises can use Gemini for business. The pages we read do not name a mainland China endpoint or a third-party API router.
How does Gemini 4 Carbon vs Claude Fable 5.2 compare on price?
Through Nov 30, 2026, Carbon's paid API rate is $2 per million input tokens and $10 per million output tokens, then $4 and $20. Anthropic lists Claude Fable 5.2 at $8 and $40. The introductory Carbon rate is one quarter of that Fable list price per token. Google also says Carbon uses more tokens than Gemini 4 Argon on hard tasks, so the token rate is not a cost-per-task quote.
Did you test Gemini 4 Carbon?
No. This page is based on official documentation, benchmarks and public reports. Google's rank claims are Google's. Where Google did not print a score, this page does not invent one.