Direct answer: Gemini 4 Carbon is now available on the paid Gemini API and in Google AI Studio under the model ID gemini-4-carbon. It is designed for difficult, long-running coding, multimodal research, scientific reasoning and defensive cybersecurity tasks. Introductory pricing is $2 per 1 million input tokens and $10 per 1 million output tokens through November 30, 2026; on December 1, the listed rates rise to $4 input / $20 output per 1 million tokens.[1][3][5]
Gemini 4 Carbon quick reference
| Item | Confirmed detail |
|---|---|
| API model ID | gemini-4-carbon |
| Availability | Paid Gemini API and Google AI Studio |
| Free tier | Not available |
| Default thinking level | medium |
| Other thinking levels | low, high |
| Context window | Unified 1,048,576-token window shared by input, thinking and output |
| Standard synchronous output limit | 65,536 tokens |
| Longer output | Background execution, up to the context limit minus input |
| Introductory input price | $2 per 1M tokens through Nov. 30, 2026 |
| Introductory output price | $10 per 1M tokens, including thinking tokens |
| Price from Dec. 1, 2026 | $4 input / $20 output per 1M tokens |
| Cached input | Implicit caching only; 95% discount on cache hits |
| Best fit | Hard debugging, architecture, dense multimodal synthesis, scientific reasoning, security auditing |
Google describes Carbon as its strongest reasoning model and says it spends more thinking tokens than Gemini 4 Argon to explore hypotheses, verify work and check edge cases. That extra reasoning can improve difficult work, but it can also increase latency and total token usage.[3]
Should you use Carbon or Argon?
Use Gemini 4 Carbon when the task is expensive to get wrong or genuinely requires extended reasoning:
- tracing a bug across several packages or services;
- reviewing an architecture or migration plan;
- synthesizing dense PDFs, spreadsheets and slides into a strict schema;
- checking a proposed patch against tests and edge cases;
- formal, scientific or mathematical reasoning;
- defensive security review within an authorized scope.
Keep Gemini 4 Argon for most interactive coding, routine feature work and faster production loops. Google says Carbon and Argon share the same listed per-token price and 1M-token unified context, but Carbon normally consumes more thinking tokens per turn. The lower-cost decision is therefore not just the price table: it is the total tokens and time each model uses to finish your workload.[3]
A practical rule is to start with Argon or a Flash model, then escalate only the hardest cases to Carbon. If Carbon is necessary, test low, medium and high, and keep the lowest thinking level that passes your quality checks.
Python quickstart
Install or update Google’s SDK, set your API key in the environment, and call the Interactions API. Do not paste a production key directly into source code.
pip install -U google-genai
export GEMINI_API_KEY="your-key-here"
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-carbon",
input=(
"Review this service design for race conditions. "
"List the failure modes, propose the smallest safe patch, "
"and provide tests that would fail before the patch."
),
generation_config={
"thinking_level": "medium"
},
)
print(interaction.output_text)
print(interaction.usage)
Google recommends medium for most complex coding and agentic work. Use low for simpler or cost-sensitive jobs, and reserve high for the hardest code, math, science or agent tasks.[3]
Five-step migration checklist
1. Change the model ID
If you already use Gemini 4 Argon, replace gemini-4-argon with gemini-4-carbon. Google says this move requires no request-schema change.[3]
model="gemini-4-carbon"
Moving from Gemini 3 or earlier needs more review because Carbon handles thinking, caching and generation settings differently.
2. Replace thinking budgets with thinking levels
Use thinking_level with low, medium or high. Do not carry forward an old thinking_budget setting. Test all three levels on representative tasks instead of assuming that high is automatically best for every request.[3]
generation_config={
"thinking_level": "medium"
}
3. Review output limits and timeouts
Carbon’s thinking tokens count against max_output_tokens. If an existing application uses a limit below 32,768, Google recommends considering at least 32,768 because deeper reasoning can consume more of the allowance before the final answer is produced. Hard prompts may also take several minutes, so use background execution or set a sufficiently long client timeout.[3]
Do not raise limits blindly. A high ceiling controls the maximum possible consumption, not the likely answer quality. Measure completion rate, useful output and cost on your own tasks.
4. Remove unsupported request settings
Google’s migration guidance says Carbon does not support explicit context caching. Remove calls that create or reference cachedContents or cached_content. Implicit caching is enabled automatically for prompts of at least 4,096 tokens, and successful cache hits receive the listed cached-input discount.[3]
Also remove temperature, top_p and top_k; Google says these sampling parameters are ignored for Carbon.[3]
5. Test quality, latency and total cost before switching production
Build a small evaluation set from real work: one routine task, one difficult task, one long-context task and one known failure case. Run the same inputs through your current model and Carbon at each thinking level. Record:
- whether the output passes an objective check;
- input, thinking and output token usage;
- time to first useful result;
- number of retries or corrective prompts;
- total cost per completed task.
Do not promote Carbon because of a benchmark headline alone. Promote it when it finishes your specific hard tasks more reliably or reduces expensive human rework enough to justify the extra token use.
Cost examples
The following examples use Google’s introductory standard rates through November 30, 2026. Output billing includes thinking tokens.[3]
Example 1: medium code review
A request uses 100,000 input tokens and 20,000 billed output-plus-thinking tokens:
- Input: 0.1 × $2 = $0.20
- Output: 0.02 × $10 = $0.20
- Estimated request cost: $0.40
At the rates listed from December 1, the same token usage would be $0.80.
Example 2: large repository analysis
A job uses 750,000 input tokens and 100,000 output-plus-thinking tokens:
- Input: 0.75 × $2 = $1.50
- Output: 0.1 × $10 = $1.00
- Estimated request cost: $2.50
The same usage at the December rates would be $5.00. These examples exclude paid grounding queries and assume no cache hit.
Reusable formula
cost = (input_tokens / 1,000,000 × input_rate)
+ (output_and_thinking_tokens / 1,000,000 × output_rate)
When grounding or tools are enabled, include their billable usage as well. Google’s documentation says internal tool instructions and function declarations processed by the model are billed, and Google Search grounding remains billed per query.[3]
How to run outputs longer than 64K tokens
Synchronous responses stop at 65,536 tokens. For a longer result, Google recommends server-managed background execution with background=True and a max_output_tokens value above 65,536. The server continues the job in segments, and your client polls until the interaction is complete.[3]
import time
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-carbon",
input="Analyze the supplied project and produce a tested migration plan.",
background=True,
generation_config={
"max_output_tokens": 131072,
"thinking_level": "high",
},
)
while interaction.status in ("queued", "in_progress"):
time.sleep(5)
interaction = client.interactions.get(id=interaction.id)
text = "".join(
item.text
for step in interaction.steps
if step.type == "model_output"
for item in step.content
if item.type == "text"
)
print(text)
print(interaction.status)
print(interaction.usage)
Set a deliberate upper limit: thinking and visible output share the allowance, and all processed output tokens are billable. Google also notes that setting background=True without a larger max_output_tokens still stops the job at 64K.[3]
Production rollout checklist
Before routing live traffic to Carbon:
- confirm the account is on the paid Gemini API tier;
- update the SDK and test the exact endpoint you will deploy;
- keep API keys in a secret manager or environment variable;
- remove explicit caching and ignored sampling fields;
- set a per-request token ceiling and application budget alert;
- log usage, latency and finish status without logging sensitive prompts;
- use background execution for long-running tasks;
- keep Argon or a Flash model as the default for routine work;
- require human review for high-impact security, legal, financial or scientific decisions;
- run only authorized cybersecurity testing against systems you own or have permission to assess.
Carbon is a reasoning tool, not proof that an answer or patch is correct. Keep tests, source validation, access controls and human approval in the workflow.
Frequently asked questions
Is Gemini 4 Carbon available now?
Yes. Google announced general availability on November 10, 2026, and lists it for paying developers through the Gemini API and Google AI Studio. It is also rolling out through selected Google products and subscriptions.[1][5]
What is the Gemini 4 Carbon API model ID?
The model ID is gemini-4-carbon.[1][3]
Is there a free API tier?
No. Google’s documentation says Gemini 4 Carbon is not available on the free tier.[3]
How much does Gemini 4 Carbon cost?
Through November 30, 2026, Google’s listed introductory standard rate is $2 per 1M input tokens and $10 per 1M output tokens, including thinking tokens. From December 1, the listed rates are $4 input and $20 output per 1M tokens.[3][5]
Does Carbon have a 1M-token context window?
Yes. Google documents a unified 1,048,576-token window shared by input, thinking and output. A very large input leaves less room for reasoning and generated output.[3]
Is Carbon always better than Argon?
No. Google positions Argon as the efficient choice for most advanced software engineering and knowledge work, while Carbon is intended for tasks that need deeper reasoning and verification. Carbon normally uses more thinking tokens and can be slower.[3][5]
Can I migrate by changing only the model name?
From Gemini 4 Argon, the request schema can remain the same, but you should still review timeouts, output limits, latency and cost. From Gemini 3 or earlier, also review thinking settings, explicit caching, sampling parameters and multi-turn state.[3]
Bottom line
Gemini 4 Carbon is a strong candidate for the small share of work where deeper reasoning and verification matter more than fast, cheap responses. Use the exact model ID gemini-4-carbon, begin with medium thinking, cap token use, and evaluate cost per successful task. Keep a faster model for normal traffic and route only difficult, high-value work to Carbon.
Sources
[1] https://aistudio.google.com/docs/changelog — Gemini API release notes
[3] https://aistudio.google.com/docs/whats-new-gemini-4-carbon — What is new in Gemini 4 Carbon
[5] https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available — Gemini 4 Carbon is now available