Gemini 4 Carbon is Google’s strongest Gemini 4 model, built on Argon and aimed at harder, longer coding, research and agent work. Google opened it to developers, enterprises, scientists and consumers on November 10, 2026, after a first stage limited to cyber defenders. Access still depends on your plan or API account.
If you already use or are evaluating Gemini 4 Argon, the useful question is not which model scores higher. Both list at the same per-token price through November 30, 2026, so the difference shows up in tokens spent per task, latency and who can log in.
This guide is based on pages Google published, read on November 18, 2026. It is not a hands-on test, benchmark numbers are Google’s own, and anything Google has not confirmed is marked as such.
What is Gemini 4 Carbon?
Gemini 4 Carbon, model ID gemini-4-carbon, is Google’s most capable Gemini 4 reasoning model. Google’s model card says it is built on Argon for tasks that test the limits of a model’s intelligence, such as multi-step coding, debugging and open-ended scientific research. It reads text, images, audio and video and returns text only.
The model card lists a knowledge cutoff of March 2026. It warns of general model limits: hallucinations, occasional slowness or timeouts, and heavier token use at higher thinking levels.
Google’s Carbon developer notes describe a unified 1,048,576-token window shared by input, thinking and output. Standard responses stop at 65,536 tokens. With background mode and a higher max_output_tokens, output can reach about 1M tokens minus the input, generated in 64K continuation hops.
When did Gemini 4 Carbon launch, and why do early reports disagree?
Google first described Carbon on October 28, 2026, as available to a subset of cyber defenders, with access expanding through its Fairwind Program. It then opened Carbon more widely on November 10, 2026. Articles that call Carbon restricted describe the first stage, not the current one.
The November 10 announcement says Google is making Carbon available to developers, enterprises, scientists and consumers. The same post introduces a preview of /swarm in Antigravity, covered below. Pages that list only a release date skip these access details. That is where most confusion starts.
Who can use Gemini 4 Carbon today?
As of November 18, 2026, Carbon is open to paid Gemini API developers, enterprises through Gemini for business, Google AI Ultra subscribers, approved scientists and vetted cyber defenders. It is not on the Free API tier. Google’s Antigravity page lists it as unavailable on Google AI Pro, Plus and Free.
| Route | Who gets Carbon | Catch |
|---|---|---|
| Gemini API, AI Studio, Android Studio | Developers on the paid tier | Google’s developer notes say Carbon is not available on the Free tier |
| Gemini for business, Antigravity, CodeMender | Enterprises | Contract terms and admin settings are not published on the pages read |
| Gemini app, AI Mode in Search, Antigravity | Google AI Ultra subscribers | Google’s announcement names Ultra only for consumer products |
| Scientists program | Approved scientists | Free six-month access to Argon and Carbon through Antigravity’s highest usage tier, subject to eligibility, capacity and selection |
| Fairwind Program | Verified cybersecurity defenders | Application and vetting, on a rolling basis |
| Google AI Pro, Plus and Free plans | Not Carbon in Antigravity | Google’s Antigravity models page lists Argon for Pro but Carbon only for Ultra and Enterprise |
Rows one to five come from Google’s November 10 announcement and developer notes. The Antigravity page also lists Argon for Pro. That differs from the Antigravity row in our guide on how to access Gemini 4 Argon, written in October. Confirm in your own account before you pay for a plan.
How much does Gemini 4 Carbon cost?
Carbon’s Standard API price is $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 from December 1. Output includes thinking tokens. Cached input is $0.10 per million, rising to $0.20. These per-token rates match Argon’s Standard rates.
Google’s developer notes add that implicit caching is on by default for prompts of at least 4,096 tokens, with a 95% discount on cache hits, and that explicit caching is not supported. All processed tokens are billed, including thinking tokens and the tool instructions used with grounding.
Not confirmed as of November 18, 2026: Carbon’s Batch, Flex and Priority rates. The pricing page I opened listed only Argon rows, so I do not state Carbon rates for those tiers. Google’s footnote also says the December rate applies to the Carbon model intended for public access, and it publishes no separate price for the Fairwind version.
Equal prices per token do not mean equal prices per task. Carbon spends more thinking tokens per turn, and Google publishes no multiplier. The table below is an illustration, not a measurement. It assumes a 30,000-token input, an Argon baseline of 10,000 output tokens, and hypothetical multipliers for Carbon. For real figures, see the Gemini 4 Argon cost per task breakdown.
| Scenario (assumed) | Output tokens | Cost per task to Nov 30 | Cost per task from Dec 1 |
|---|---|---|---|
| Argon baseline | 10,000 | $0.16 | $0.32 |
| Carbon at 1.5x output | 15,000 | $0.21 | $0.42 |
| Carbon at 2x output | 20,000 | $0.26 | $0.52 |
| Carbon at 3x output | 30,000 | $0.36 | $0.72 |
The arithmetic ignores caching and grounding. Even at 3x, a task costs under a dollar. Carbon’s extra spend matters at volume, not per request. At 2,000 tasks a month, the 3x row adds roughly $400 over the Argon baseline before December and $800 from December 1.
Measure Your Own Multiplier
Run the same five real tasks on both models at medium thinking, record billed input and output tokens from the response metadata, and divide Carbon’s output by Argon’s. That one ratio replaces every assumption in the table above.
How does Gemini 4 Carbon differ from Argon?
According to Google’s documentation, Carbon and Argon share the 1M context window, built-in tools and per-token prices. The main difference is that Carbon spends more thinking tokens per turn to explore thoroughly and verify its work, which raises token use and latency. Argon is the faster, lower-spend option for interactive work.
| Item | Gemini 4 Argon | Gemini 4 Carbon |
|---|---|---|
| Model ID | gemini-4-argon | gemini-4-carbon |
| Context window | 1,048,576 tokens, shared | 1,048,576 tokens, shared |
| Thinking levels | low, medium (default), high | low, medium (default), high |
| Standard price to Nov 30 | $2 in, $10 out | $2 in, $10 out |
| Standard price from Dec 1 | $4 in, $20 out | $4 in, $20 out |
| Free API tier | No | No |
| Antigravity plans | Ultra, Enterprise, Pro | Ultra, Enterprise |
| Google’s positioning | Faster responses, lower token spend per turn | Strongest reasoning, sustained agentic workflows |
Google’s model card gives selected benchmark pairs, with Argon in parentheses: Artificial Analysis Intelligence Index v5.0 at 59 (50), GraphWalks from 256k to 1M tokens at 96.2% (84.2%), and ExploitGym at 54.8% (42.0%). The same card shows Carbon trailing Claude Opus 5.5 on an offline subset of OSWorld-2.1, 81.5% against 87.2%. All of these are Google-reported. For rival models, see Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5.
How hard is it to switch from Argon to Carbon?
Google says to change the model ID from gemini-4-argon to gemini-4-carbon, with no request schema changes. Keep thinking_level at medium unless a workload needs high, and make sure max_output_tokens and HTTP timeouts allow longer reasoning. Google notes hard prompts can take several minutes and suggests background execution or timeouts of at least 15 minutes.
The developer notes also say temperature, top_p and top_k are ignored if provided, so remove them rather than assume they tune anything.
Argon or Carbon: which should L&D and product teams use?
Start on Argon for most work. Google’s announcement says Argon “remains the ideal choice for the majority of software engineering and knowledge work,” and positions Carbon for harder tasks that use more compute. Move a task to Carbon only when Argon fails your acceptance check and the extra tokens are worth the cost.
The table below is our own decision aid. It uses only the published facts above and labels each judgment that goes beyond them.
| Situation | Start with | Published basis | Assumption to test |
|---|---|---|---|
| Interactive coding, daily feature work, fast production loops | Argon | Google describes Argon as faster with lower token spend per turn | Your users notice latency |
| Hard debugging, long multi-step engineering, sustained agent runs | Carbon | Google says Carbon is built for sustained agentic workflows | Argon’s failures cost more than Carbon’s extra tokens |
| Long source sets for course or policy drafts | Argon, then Carbon for failures | Same 1M window; Carbon scores higher on GraphWalks in Google’s card | Your documents resemble that test, which is unproven |
| Compliance or policy comparison with SME review | Argon, then Carbon if the SME queue is too large | Google reports 92% against 70% for Argon on its own 375-task honesty benchmark | That benchmark transfers to training content; vendor-reported, not independent |
| Security review and patching | Carbon through Fairwind | Google ties Carbon’s cyber access to that program | Your team qualifies as a verified defender |
| Routine, high-volume tasks such as tagging or summaries | Test Hydrogen first | Google lists Hydrogen at $1.50 in and $7.50 out per million tokens | Quality is good enough at that tier |
| Team on Google AI Pro, no Ultra | Argon or Hydrogen | Google’s Antigravity page does not list Carbon for Pro | Your Pro plan still shows these models |
For compliance and safety content, keep a subject matter expert in the loop with either model. Both documents list hallucination as a known limit, and a higher benchmark score does not remove the need for review.
What does /swarm do, and does it need Gemini 4 Carbon?
/swarm is an Antigravity command, in preview, that turns one prompt into a coordinated team of agents you can track in real time. It is not Carbon-only. Google’s swarm documentation says it runs on Google AI Pro and Ultra plans, while Carbon in Antigravity is listed for Ultra and Enterprise.
Per Google’s swarm documentation, your current conversation becomes the coordinator. It splits the task, adds named workers and brings in a verifier to read diffs and run tests. Members talk on a shared message board, and Google warns that a swarm uses quota much faster than a single conversation.
The November 10 post suggests assigning Carbon to demanding roles such as architecture, coordination and auditing, and pairing it with faster models. Our inference, not Google’s statement: on a Pro plan a swarm would run without Carbon.
Test Swarm Small First
Google says swarms work best when a task splits into pieces that do not collide, such as one worker per file or package. Try one on a small repository, then check how much quota the run consumed before pointing it at real work.
Where does Gemini 4 Hydrogen fit beside Carbon and Argon?
Gemini 4 Hydrogen is a third model, announced by Google on November 12, 2026, two days after Carbon. Google calls it a fast, cost-efficient model for everyday coding and agentic workflows, with model ID gemini-4-hydrogen and prices of $1.50 per million input tokens and $7.50 per million output tokens.
In Google’s framing, Carbon coordinates and Hydrogen executes in multi-model workflows. Google lists Hydrogen on its Free plan and in the free API tier, which Carbon and Argon do not have. This post does not compare Hydrogen further, since Google’s figures for it are also vendor-reported.
What is still not confirmed about Gemini 4 Carbon?
Several details are missing from Google’s pages as of November 18, 2026. Treat these as open until Google publishes them or you verify them in your account:
- Carbon’s Batch, Flex and Priority prices, and the price of the Fairwind version.
- Whether Google AI Pro or Plus subscribers get Carbon in the Gemini app, since the announcement names only Ultra.
- Rate limits, regional availability and enterprise terms for Gemini for business.
- Independent confirmation of Google’s benchmark figures, including its cost-per-task claims against other labs’ models.
What should you do next?
Put November 30, 2026 on your calendar. Standard rates double on December 1, so any budget built on the current $2 and $10 prices needs a second column. Our Gemini 4 Argon pricing guide shows the tier-by-tier changes.
Then pick five real tasks from your own work, such as a long policy summary, a quiz draft from an SME document and a code review. Run each on Argon and Carbon at medium thinking in AI Studio. Record accepted outputs, billed tokens and time. If Carbon does not win on accepted results per dollar for a task type, keep that task on Argon.
FAQ
Q1. What is Gemini 4 Carbon?
Gemini 4 Carbon is Google’s strongest Gemini 4 reasoning model, built on Gemini 4 Argon. Google’s model card describes it as made for tasks that test a model’s limits, such as multi-step coding, debugging and scientific research. It accepts text, images, audio and video and returns text only. Google opened it more widely on November 10, 2026.
Q2. Is Gemini 4 Carbon available to everyone?
No. As of November 18, 2026, Google names paid Gemini API developers, enterprises using Gemini for business, Google AI Ultra subscribers, approved scientists and verified cyber defenders in the Fairwind Program. Google’s Antigravity page lists Carbon as unavailable on Google AI Pro, Plus and Free plans. The API has no Free tier for it.
Q3. How much does Gemini 4 Carbon cost?
Standard API pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1 it becomes $4 and $20. Output includes thinking tokens, and cached input is $0.10 per million, rising to $0.20. Carbon’s Batch, Flex and Priority prices were not confirmed on Google’s pricing page as of November 18.
Q4. What is the Gemini 4 Carbon model ID and context window?
The model ID is gemini-4-carbon. Google’s developer notes describe a unified 1,048,576-token window shared by input, thinking and output tokens. Standard responses stop at 65,536 tokens. Background mode with a higher max_output_tokens can reach about 1M tokens minus the input, generated in 64K continuation hops.
Q5. Is Gemini 4 Carbon better than Argon?
On Google’s published benchmarks, Carbon scores higher, for example 59 against 50 on the Artificial Analysis Intelligence Index v5.0. It also spends more thinking tokens per turn, which raises token use and latency. Google says Argon remains the ideal choice for most software engineering and knowledge work, so Carbon suits the harder tasks.
Q6. What is /swarm in Google Antigravity?
/swarm is an Antigravity command, in preview, that turns one prompt into a coordinated team of agents you can track in real time. A coordinator splits the task, workers handle the pieces and a verifier reviews changes. Google warns that a swarm uses quota much faster than a single conversation.
Q7. Does Gemini 4 Carbon have a free API tier?
No. Google’s developer notes state that Gemini 4 Carbon is not available on the Free tier, so you need a billing-enabled Gemini API project. Gemini 4 Hydrogen, a separate and cheaper Google model announced November 12, 2026, is listed on the free tier in Google’s API changelog.