On November 10, 2026 (US time), Google made Gemini 4 Carbon generally available. The model was announced on October 28, and for its first two weeks only a limited set of cyber defense organizations could use it. Google calls Carbon its "most capable model" and places it above Gemini 4 Argon, which became generally available on October 13.
The three most common questions in search are "How is it different from Argon?", "How much does it cost?", and "Can I use it with Google AI Pro?" The short answers: the API price per token and the context length are the same as Argon; what differs is how much thinking goes into each response. In the Gemini app and Antigravity, only Google AI Ultra subscribers can use it; it isn't available on Pro. In the API it's paid tier only, and the price is $2 input and $10 output until November 30, rising to $4 and $20 on December 1 (checked November 11, 2026, against Google's announcement, the Gemini API documentation, the model card, and the Google Cloud documentation).
Where you can use Gemini 4 Carbon (checked November 11, 2026)
Gemini API and Google AI Studio (paid tier)Not on the free tier (Ultra subscribers can also try it in AI Studio)
AvailableGemini app and AI Mode in SearchRolling out to Google AI Ultra (Ultra 5x and Ultra 20x) subscribers
Ultra onlyGoogle AI Pro, Plus, and the free appThe top model on Pro is still Argon
Not availableGoogle AntigravityFor Ultra subscribers (/swarm, which runs multiple agents, is in preview; researchers can also apply for a free six-month allowance)
Ultra onlyGoogle Cloud (Gemini Enterprise Agent Platform)For approved enterprise customers, in some regions
Approval requiredGitHub CopilotPro+, Max, Business, and Enterprise (pay for what you use at Google's list price; rolling out)
Available
Sources: Gemini 4 Carbon is now available (Google DeepMind, November 10, 2026); Gemini Developer API pricing; Gemini plans (prices vary by country); Gemini 4 Carbon page on Google Cloud; GitHub Changelog (all checked November 11, 2026).
1. What is Gemini 4 Carbon? A model that "thinks more," placed above Argon
As Google describes it, Argon is the standard model for most software development and knowledge work, while Carbon is for work that tests the limits of model intelligence, such as hard multi-step coding and debugging and open-ended research. Carbon uses more tokens than Argon to explore hypotheses more broadly, reason more deeply, and verify for longer (What's new in Gemini 4 Carbon).
The names are a separate series from the Flash and Pro names Gemini has used so far. Gemini 4 has two models, Argon and Carbon, and the Gemini API's model list shows Carbon first and Argon second. For the names in the Gemini app and how they relate to other models, see What is Google Gemini.
| Item | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Model ID | gemini-4-carbon | gemini-4-argon |
| Release | Announced October 28, 2026; generally available November 10 | Announced September 30, 2026; generally available October 13 |
| Context length | 1,048,576 tokens for input, thinking, and output combined (same) | |
| Output per call | Normally up to 65,536 tokens; with "long output" in background execution, close to 1 million tokens minus the input | |
| Thinking levels | low, medium (default), high | |
| Inputs | Text, images, audio, video, PDF (output is text only) | |
| Knowledge cutoff | March 2026 | March 2026 |
| Thinking per response | More (slower, more tokens per response) | Less (faster, lower cost per response) |
Sources: Gemini 4 Carbon model page; Gemini 4 Carbon model card; What's new in Gemini 4 Carbon (checked November 11, 2026). Argon's release dates and cutoff are from the Argon model card. Cutoffs for models from each company are collected in AI knowledge cutoff dates.
2. The biggest difference from Argon is how much it thinks per response
Google's documentation describes the main difference between the two models this way: context length, available tools, and price per token are the same; what mainly differs is the number of thinking tokens used in each response. Because Carbon explores and verifies more thoroughly, it uses more tokens and takes longer per response than Argon.
Same as Argon
- Price per token ($2 and $10 until November 30; $4 and $20 from December 1)
- The 1-million-token window and long output
- Tools such as Grounding with Google Search, code execution, function calling, and computer use
- Three thinking levels, with
mediumas the default
What changes with Carbon
- It thinks more per response, so output tokens (including thinking) increase
- It takes longer before the first token appears (several minutes on hard questions)
- Accuracy on hard tasks improves (in Google's table the gap is largest on long context; the cyber gap uses the vetted version's score)
Even at the same price per token, Carbon uses more tokens per response. The output price includes thinking tokens, and Carbon uses more of them. In third-party measurements by Artificial Analysis, Carbon produced 205 million tokens on the same evaluation tasks versus 113 million for Argon. In the same measurements, however, Carbon's total cost per task came out about 20% lower (Section 5). Which one ends up cheaper depends on the work, so measure both on your own tasks before deciding.
3. Pricing: launch pricing until November 30, doubling on December 1
Carbon's API price starts at the same launch pricing as Argon and doubles from December 1, 2026. Argon rises to the same price on the same day.
Until November 30, 2026
$2 / $10
Input / output per million tokens
From December 1, 2026
$4 / $20
Same price as Claude Opus 5.5
| Tier | Input | Output (including thinking) | Cache reads |
|---|---|---|---|
| StandardNo free tier | $2.00 → $4.00 | $10.00 → $20.00 | $0.10 → $0.20 |
| Batch and FlexNon-urgent processing | $1.00 → $2.00 | $5.00 → $10.00 | $0.05 → $0.10 |
| PriorityPrioritized processing | $3.60 → $7.20 | $18.00 → $36.00 | $0.18 → $0.36 |
Per million tokens, in US dollars. Left: until November 30; right: from December 1. Source: Gemini Developer API pricing (checked November 11, 2026). Grounding with Google Search is free for up to 5,000 queries a month across all Gemini 3 and later models, then $14 per 1,000 queries (up to 1,500 a day).
Caching is automatic caching (implicit caching) only: it applies automatically to repeated prefixes of 4,096 tokens or more, and input that hits the cache is 95% off. Creating caches yourself (explicit caching) isn't supported. When you use tools such as Google Search, the tokens for instructions and function definitions that Google adds behind the scenes are also billed.
Compared with top models from other companies
| Model | Input | Output | Notes |
|---|---|---|---|
| Gemini 4 Carbon | $2 → $4 | $10 → $20 | Right-hand price from December 1 |
| Gemini 4 Argon | $2 → $4 | $10 → $20 | Same as Carbon |
| Claude Opus 5.5 | $4 | $20 | Same as Carbon from December 1 |
| Claude Fable 5.2 | $8 | $40 | Twice Carbon's price from December |
| GPT-6 Astra | $10 | $50 | For short context |
Per million tokens, in US dollars (each company's price list as of November 11, 2026). The differences between the two Claude models are covered in Claude Fable 5.2 vs. Opus 5.5, and Gemini's free version, plans, and API pricing as a whole in Gemini pricing and what you can do for free.
4. Benchmarks: Google's comparison table
The model card has a table comparing Carbon with Argon, GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1. These are Google's own numbers, and Claude Fable 5.2, released November 9, isn't included. Here are the rows with the largest and smallest gaps.
| Evaluation | Carbon | Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| GraphWalksLong context (256K–1M) | 96.2% | 84.2% | 71.8% | 66.8% |
| ExploitGymVulnerability exploitation (Carbon's score is from the vetted version) | 54.8% | 42.0% | 42.4% | 34.5% |
| Terminal-Bench 4.0 (AA)Work in the terminal | 65.2% | 57.1% | 59.1% | 59.6% |
| AutomationBenchWork automation (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% |
| FrontierSWE 2Software development | 61.2% | 55.0% | 65.5% | 62.3% |
| OSWorld-2.1Computer use | 81.5% | 74.7% | 80.9% | 87.2% |
| ChartographyReading charts | 72.5% | 71.6% | 71.0% | 66.3% |
Source: Gemini 4 Carbon model card (checked November 11, 2026). Scores are from evaluations and conditions chosen by Google, which differ from other companies' announcements. Rows marked "(AA)" were measured by Artificial Analysis. Carbon's score on the cyber evaluation is from use through the Fairwind Program for vetted organizations. The full table (including the Claude Fable 5.1 column) is in the model card.
The gap with Argon is large on long context (12 points on GraphWalks) and 1 to 3 points on reading charts and work automation. On the cyber evaluation ExploitGym, Carbon is also 13 points ahead, but that is the score of the vetted version with adjusted safeguards. The generally available Carbon applies safeguards to output that could be used for attacks, so the same gap won't necessarily appear (Section 8). GPT-6 Astra and Claude Opus 5.5 scored higher on the software development benchmark FrontierSWE 2, and Claude Opus 5.5 scored higher on the computer-use benchmark OSWorld. Carbon isn't "the best at everything."
5. Third-party measurements: intelligence score, and cost and time per task
Artificial Analysis, which measures models independently, publishes a composite score that combines several evaluations (Intelligence Index v5.0.1). The values compare each model at its highest thinking setting.
Source: Artificial Analysis model pages (Gemini 4 Carbon, Gemini 4 Argon, Claude Fable 5.2, Claude Opus 5.5, GPT-6 Astra; checked November 11, 2026). Costs are calculated at Carbon's and Argon's prices from December 1 ($4 and $20). Values change when models are re-measured.
Three things stand out. First, Carbon's composite score is 10 points higher than Argon's, level with Claude Opus 5.5, and 5 points short of Claude Fable 5.2. Second, although Carbon produced about 1.8 times as many tokens as Argon, its cost per task was about 20% lower than Argon's, and it took less time. The cost also includes input and cache reads, but Artificial Analysis doesn't explain why Carbon came out cheaper. Third, in the same measurements, the wait before the first token was about 68 seconds for Carbon and about 40 seconds for Argon. That is far shorter than the highest thinking settings of Claude Opus 5.5 (about 358 seconds), Claude Fable 5.2 (about 230 seconds), and GPT-6 Astra (about 246 seconds), but compared with Argon, it isn't suited to uses where you want quick replies in conversation.
6. When to use Carbon and when to use Argon
Google says "Argon suits most work, and Carbon is for work that tests the limits." If you're unsure, the safe choice in terms of wait time is to start with Argon and move up to Carbon only when Argon falls short on your work. Carbon can sometimes be cheaper in total cost per task, so try both and compare.
- Carbon
Root-cause investigation and design reviews in large repositories Migrations that span multiple packages, or tracking down bugs whose cause is unknown. Google cites examples of Carbon working on a task for days.
- Carbon
Having it read long documents in full When searching for answers in documents of hundreds of thousands of tokens, the GraphWalks gap (96.2% vs. 84.2%) matters.
- Carbon
Investigating and fixing vulnerabilities Defensive verification and writing fix patches. The general Carbon applies safeguards to output that could be used for attacks, and the high cyber scores are from the vetted version (Section 8).
- Argon
Interactive coding and everyday feature work The uses Google recommends Argon for. Replies are faster, and the cost per response is lower.
- Argon
Using the app or Antigravity on Google AI Pro Carbon can't be selected on Pro, so Argon is the top option.
7. What to watch for when switching to Carbon in the API
Switching from Argon
If you already use Argon, you don't need to change the shape of your requests. Change the model ID to gemini-4-carbon and review the settings affected by the longer thinking time.
- Change the model ID from
gemini-4-argontogemini-4-carbon. - Try
thinking_levelat the defaultmediumfirst. Usehighonly for genuinely hard tasks. - Increase
max_output_tokensand the HTTP timeout. Google says hard questions can take several minutes and recommends a timeout of 15 minutes or more, or using background execution. - Keep latency-sensitive processing on Argon.
Switching from Gemini 3 or earlier
If you point Gemini 3 code at Carbon as is, some settings won't work or will have no effect.
| Current setting | On Carbon | How to fix |
|---|---|---|
thinking_budget, or minimal for thinking_level | HTTP 400 error (per Google Cloud's documentation) | Use low, medium, or high |
Explicit caching (cachedContents, caches.create) | Not supported | Remove the calls that create or reference caches and rely on automatic caching |
temperature, top_p, top_k | Ignored | Remove them and control output variation through instructions |
max_output_tokens below 32K | Thinking can use it up, so answers are easily cut short | Set it to 32,768 or more (to use fewer tokens, lower the thinking level) |
| You want more than 64K of output | Normal calls stop at 65,536 | Set background to true and max_output_tokens above 65,536 |
Sources: Migration checklist in What's new in Gemini 4 Carbon; Gemini 4 Carbon page on Google Cloud (checked November 11, 2026). Problems caused by the automatic redirection of Gemini 3 model names are covered in Gemini Flash and Pro auto-redirects and migration.
It can't be called on the free tier. Specifying gemini-4-carbon with a free API key won't work, so set up billing in AI Studio before trying it. The monthly Google Cloud credits included with Google AI Pro and Ultra ($10 for Pro; $40 or $100 for Ultra) can be applied to paid-tier charges (you first need to prepay $5 or more in AI Studio). How to claim them is covered in API credits included with subscriptions.
8. The app, Antigravity, and the cyber version
In the Gemini app and AI Mode in Search, Carbon has been rolling out to Google AI Ultra subscribers since November 10. On the Gemini plans page, Carbon is listed under Google AI Ultra (Ultra 5x and Ultra 20x), while the Google AI Pro column only goes up to Argon (checked November 11, 2026; prices vary by country). Carbon in AI Mode in Search is available only in some countries, listed in the notes on the same page.
Google Antigravity also offers Carbon to Ultra subscribers and has added a feature called /swarm in preview. From a single instruction it builds a team of agents, assigning roles that need deep thinking, such as design and review, to Carbon and hands-on roles to lighter models, and runs them in parallel. For Antigravity's limits by plan, see the Antigravity section of Gemini pricing and what you can do for free.
For cyber defense organizations, there is Gemini 4 Carbon for cyber, with adjusted safeguards, available only to organizations that apply to Google's Fairwind Program and pass review. The generally available Carbon applies safeguards to output that could be used for attacks. Google says that in a simulated attack environment, Carbon's safeguards stopped 99% of attack sequences. For the big picture on AI and cyberattacks and defense, see How AI is changing cybersecurity.
Summary
- Gemini 4 Carbon is Google's top model, generally available since November 10, 2026. It has the same price and context length as Argon but thinks more per response.
- The API price is $2 input and $10 output until November 30, and $4 and $20 from December 1 (the same as Argon). There is no free tier.
- In the app and Antigravity, it's Google AI Ultra only. The top model on Pro is still Argon.
- In Google's table, the gap with Argon is largest on long context (the high cyber score is from the vetted version). On some development and computer-use evaluations, Claude Opus 5.5 or GPT-6 Astra scores higher.
- Its third-party composite score is 60, level with Claude Opus 5.5 (59) and below Claude Fable 5.2 (65). Its cost per task was about 20% lower than Argon's, but the wait before the first token is longer than Argon's.
- When switching in the API, remove
thinking_budgetand explicit caching, and increase the output limit and the timeout.
FAQ
Q. Should I use Gemini 4 Carbon or Argon?
Google recommends Argon for most development and knowledge work. The safe choice in terms of wait time is to move up to Carbon only for work where you've found Argon isn't enough, such as root-cause investigation in large codebases or reading documents of hundreds of thousands of tokens.
Q. Can I use Gemini 4 Carbon with Google AI Pro?
Not in the Gemini app or Antigravity. Carbon is for Google AI Ultra subscribers, and the top model you can select on Pro is Gemini 4 Argon. In the API, if you prepay $5 or more in AI Studio to move to the paid tier, you can apply the $10 monthly Google Cloud credit included with Pro to the charges and call Carbon.
Q. Can I use Gemini 4 Carbon for free?
No. It isn't offered on the Gemini API free tier, and in the Gemini app it's only on the paid Google AI Ultra.
Q. How much does the Gemini 4 Carbon API cost?
Per million tokens, it's $2 input and $10 output until November 30, 2026, and $4 input and $20 output from December 1. Output includes thinking tokens, and Carbon thinks more per response than Argon. That said, in third-party measurements there are cases where Carbon's total per task came out cheaper.
Q. When did Gemini 4 Carbon become available?
It was announced on October 28, 2026, and at first offered only to some cyber defense organizations. General availability for developers, enterprises, and Ultra subscribers began on November 10, 2026 (US time).
Q. Will my Argon code work if I point it at Carbon as is?
Yes. You only change the model ID; the request shape is the same. Because thinking takes longer, though, increase max_output_tokens and the HTTP timeout. If your Gemini 3 or earlier code uses thinking_budget, you'll get a 400 error.
Update history
- November 11, 2026First published.
Sources
- Gemini 4 Carbon is now available (Google DeepMind, November 10, 2026)
- Gemini 4 Carbon (announcement) (Google, October 28, 2026)
- What's new in Gemini 4 Carbon (Gemini API)
- Gemini 4 Carbon model page (Gemini API)
- Gemini Developer API pricing
- Gemini API release notes (entry for November 10, 2026)
- Gemini 4 Carbon model card
- Gemini 4 Carbon (Google Cloud, Gemini Enterprise Agent Platform)
- Gemini plans (prices vary by country)
- Gemini 4 Carbon in GitHub Copilot (GitHub Changelog, November 10, 2026)
- Gemini 4 Carbon (Artificial Analysis)
