Gemini 4 Carbon vs Argon: Which Should You Pay For Before 30 November?

Gemini 4 Carbon vs Argon: which should you use?

Google says 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, such as complex multi-step coding and debugging and open-ended scientific research ^[1]. Carbon also uses more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops ^[1].

Our suggestion, not Google's: start a task on Argon, and move it to Carbon only if Argon falls short.

What changed on 10 November 2026?

At the end of October, Google released Carbon to a set of cybersecurity defenders through its Fairwind Program ^[1][2]. DeepLearning.AI's report on that release described a small set of Google-selected companies and no set date for a wider release ^[3].

On 10 November, Google opened Carbon to paying developers and began rolling it out to AI Ultra subscribers ^[1]. Reports from before that date, such as DeepLearning.AI's ^[3], describe only the limited release. Check the date on anything that says builders cannot use Carbon yet.

Who can use Gemini 4 Carbon today?

Access depends on how you buy Google's AI:

• Developers and companies: Carbon is available through the API in Google AI Studio and Gemini for business for all paying developers and enterprises ^[1]. Google points developers to the Gemini API paid tier in AI Studio and Android Studio ^[2].
• Enterprises: they can also reach it through the model drop-down in Gemini for business and through agents such as CodeMender and AlphaEvolve ^[2].
• Consumers: Google AI Ultra subscribers can access Carbon in the Gemini app, AI Mode in Search and Antigravity, which Google describes as a rollout ^[1][2]. Argon stays available to AI Pro subscribers and up ^[2].
• Scientists: researchers can apply for Antigravity's highest usage tier, free for six months, subject to eligibility, capacity and selection. It follows a program Google opened for 10,000 scientists ^[1].
• Cyber defenders: the cyber version is available by applying to the Fairwind Program, on a rolling basis. Its benchmark results apply to that version, and Google does not say the public model matches them ^[1].

The sources name the Gemini API paid tier and say nothing about free-tier API use. They name AI Ultra for the consumer apps and do not say AI Pro subscribers get Carbon, so check your model picker before you plan around it. In the Gemini app, 4 Argon is labelled "Code & analysis" and 4 Carbon "Most complex work" ^[2].

How do Argon and Carbon compare side by side?

Everything in this table comes from Google's announcement and 9to5Google's report on it:

	Gemini 4 Argon	Gemini 4 Carbon
Google's pitch	The ideal choice for most software engineering and knowledge work [1]	Built for tasks that test the limits of a model's intelligence [1]
Token use	Fewer tokens than Carbon [1]	More tokens than Argon [1]
Introductory price	$2 input, $10 output per 1M tokens [1]	The same introductory price [1]
Gemini app plan	AI Pro and up [2]	AI Ultra [2]
Model picker label	Code & analysis [2]	Most complex work [2]

During the introductory period Carbon costs the same per token as Argon ^[1], so any extra cost comes from the tokens it uses. After 30 November the comparison is open, because these sources give no later price for Argon.

How much does Gemini 4 Carbon cost, and what changes on 30 November?

Carbon costs $2 per million input tokens and $10 per million output tokens, the same introductory price as Argon ^[1]. 9to5Google reports that this price runs until 30 November, and that afterwards Carbon costs $4 per million input tokens and $20 per million output tokens ^[2]. Both rates double.

Take a job that sends 1 million tokens in and gets 1 million tokens back. It costs $12 at the introductory rate and $24 afterwards.

Token price is not task price. Carbon uses more tokens than Argon ^[1], so a hard task can cost more on Carbon even at the same per-token rate.

Google's cost claim compares Carbon with the next best external model, not with Argon. Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task ^[1].

What strengths does Google claim for Carbon?

Google gives three examples. The honesty and protein results come from Google's own tests, and the finance result is on Hebbia's evaluation, from pre-release testing ^[1]:

• Honesty: on Google's 375-task agentic honesty benchmark, Carbon passed 92% of tasks, against 70% for Argon. Google says it declined to fill in missing experimental data 98% of the time ^[1].
• Science: across seven low-homolog protein families from the CASP17 blind test set, Carbon retrieved 40 times more homologs on average. Google adds that broader benchmarking is still underway ^[1].
• Finance: in pre-release evaluations, Carbon matched the best model tested on Hebbia's E2E Financial Agentic evaluation ^[1].

Only the honesty result compares Carbon with Argon directly, and Google ran it. Read all three as a reason to run your own test, not as proof.

How good is Gemini 4 Carbon, and which numbers can you trust?

Treat the launch numbers as Google's claims. The #1 rankings on CursorBench 4.0, WANDR and AutomationBench appear in Google's announcement as rankings, which Google calls independent benchmarks. They are not scores, so you cannot set them against Argon ^[1]. For the cyber version, DeepLearning.AI notes that its benchmark scores are self-reported ^[3].

Two details matter if you build with it:

• Context: DeepLearning.AI's report on the first release lists input of up to 1 million tokens ^[3]. Check the model card for the current API limit.
• Prompt injection: Google says Carbon reached a 0% attack success rate on the Gray Swan IPI benchmark. In the same post it says prompt injection robustness remains an area for continued improvement ^[1].

If your agent reads web pages or email, keep a human approval step on any action that spends money or sends data.

What should you do before 30 November?

Run your own test, because your tasks are the only benchmark that decides your bill:

1. Pick five to ten of your hardest real tasks.
2. Run each on Argon and on Carbon, and record whether it finished, the tokens used and the cost.
3. Compare cost per finished task, not cost per token. Log the model name and the date with each run, so a later price change does not blur the comparison.
4. Send only the tasks where Carbon wins to Carbon, and leave the rest on Argon.

Then set a reminder for 30 November. After that date, Carbon's rates become $4 and $20 per million tokens ^[2], so a budget built on the introductory price will run out twice as fast.