In short: Gemini 4 Carbon, released on November 10, 2026, costs $2 per million input tokens and $10 per million output tokens on the Gemini API until November 30, then the rate doubles to $4 and $20 on December 1. That is exactly the same per-token price as Gemini 4 Argon, so the real cost difference between the two is how many thinking tokens each model burns per task, not the rate card. There is no free tier for either model, so this post works through what a typical job actually costs and when paying for Carbon's extra thinking makes sense.
What Gemini 4 Carbon is, in one paragraph
Carbon (model ID gemini-4-carbon) is Google's newest and strongest Gemini model. The official "What's new" guide describes it as built for sustained agentic coding, multimodal knowledge work, scientific reasoning and cybersecurity defense. On hard problems it deliberately spends more thinking tokens to explore hypotheses and check its own work before answering. It shares a unified context window of 1,048,576 tokens with Argon (input, thinking and output all count against that one window), returns up to 65,536 tokens in a normal synchronous call, and can go up to about 1M output tokens with the new long decode option. Inputs can be text, images, video, audio and PDF; output is text.
Where you can use it: the paid tier of the Gemini API (in Google AI Studio and Android Studio), Gemini for business, and, for Google AI Ultra subscribers, the Gemini app, AI Mode in Search and Google Antigravity. This post is about the API bill, because that is the part with a published per-token price and a deadline.
The rate card, read carefully
All prices below are per 1 million tokens in US dollars, taken from the official Gemini Developer API pricing page on November 12, 2026. Carbon has four ways to buy the same model:
- Standard (normal interactive calls): $2.00 input and $10.00 output through November 30, 2026; $4.00 and $20.00 from December 1.
- Batch (asynchronous jobs): $1.00 and $5.00 through November 30; $2.00 and $10.00 from December 1.
- Flex (synchronous, best-effort, variable latency; still marked preview in the docs): the same numbers as Batch.
- Priority (higher-reliability lane): $3.60 and $18.00 through November 30; $7.20 and $36.00 from December 1.
Three lines on that page change how you should estimate cost:
1. Thinking tokens are billed as output
The output price is labeled "including thinking tokens", and the pricing note says every token the model processes is billed, including thinking tokens and internal tool instructions and function declarations when grounding is used. Since Carbon is designed to think longer than Argon, output is where your money goes. A reply that shows you 2,000 words may have cost you many times that in hidden reasoning.
2. Caching is implicit only
Carbon does not support explicit context caching (cachedContents / caches.create); the migration checklist tells you to remove those calls. Implicit caching is on by default for prompts of at least 4,096 tokens and discounts cached input by 95%, to $0.10 per million tokens until November 30 and $0.20 after. If you send the same large system prompt or repository context again and again, structure your prompts so the repeated part comes first and stays identical.
3. Grounding is a separate meter
Grounding with Google Search includes 5,000 free search requests per month, shared across all Gemini 3 and newer models, then costs $14 per 1,000 requests, with a default cap of 1,500 requests per day per project. One user prompt can trigger more than one search query, and each query is billed. Grounding with Google Maps has the same 5,000 free, then $14 per 1,000 structure.
What a job actually costs: my own math
Per-million prices are hard to feel, so I turned the official rates into four example jobs. The token counts are illustrative workloads I chose, not measurements, and the prices are straight multiplication of the published rates:
- Code review turn, 60K tokens in and 20K out (thinking included): $0.32 on Standard until November 30, $0.64 from December 1, $0.16 on Batch or Flex, $0.58 on Priority.
- Agent bug-fix run, 400K in and 120K out: $2.00 now, $4.00 after the intro period, $1.00 on Batch or Flex, $3.60 on Priority.
- Same run with a warm cache, where 300K of the 400K input tokens are implicit cache hits: $1.43 now (100K × $2 + 300K × $0.10 + 120K × $10, per million), $2.86 after December 1. Caching saves about 29% here because output dominates the bill.
- Long-decode job, a 48.6K-token prompt that fills the rest of the 1M window with output: about $10.10 on Standard now and $20.19 from December 1.
The pattern is clear in every row: output, including thinking, is 60% to 99% of the cost. Cutting input helps a little; controlling how much the model thinks helps a lot.
Carbon vs Argon vs 3.8 Flash
Because Carbon and Argon share one rate card, the useful comparison is with a cheaper tier as well. Everything in this table comes from the official pricing, model and "What's new" pages, checked today.
| Model | Free tier | Standard input / output per 1M | Batch input / output per 1M | Positioning (official) |
|---|---|---|---|---|
| Gemini 4 Carbon | Not available | $2 / $10 through Nov 30, 2026; $4 / $20 from Dec 1 | $1 / $5, then $2 / $10 | Strongest reasoning; spends more thinking tokens per turn; 1M unified context |
| Gemini 4 Argon | Not available | $2 / $10 through Nov 30, 2026; $4 / $20 from Dec 1 | $1 / $5, then $2 / $10 | Efficient Gemini 4 model; faster, lower token spend per turn; same 1M context. Invite-only Contributor tier at $1 / $5 |
| Gemini 3.8 Flash | Yes (free input and output, content used to improve products) | $0.75 / $3.75 through Dec 31, 2026; $1.50 / $7.50 from Jan 1, 2027 | $0.375 / $1.875, then $0.75 / $3.75 | Google's "most intelligent Flash model", aimed at long-horizon coding and agents |
Two things stand out. First, Argon's intro rate also ends November 30, so "wait and use Argon instead" does not avoid the December price jump. Second, 3.8 Flash's intro rate runs a month longer, to December 31, and it is the only one of the three you can try on the free tier.
When Carbon is worth it over Argon
Google's own guidance draws the line by reasoning depth. Use Carbon for complex root-cause debugging, architectural design, dense multimodal synthesis (many PDFs, spreadsheets, drawings), scientific proofs and security auditing. Keep Argon for interactive coding, everyday feature work and fast production loops, where it answers faster and spends fewer tokens per turn.
Turning that into a budget rule: since the rates are identical, Carbon costs more on a task only in proportion to the extra tokens it uses. If Carbon takes twice the output tokens of Argon on a bug-fix run, it costs roughly twice as much; that is worth it only when Argon's cheaper attempt would fail and need a second or third pass, or a human hour. A practical approach:
- Pick 10 to 20 real tasks from your backlog, including a few hard ones Argon struggled with.
- Run them on both models at the default
thinking_levelofmedium, and log input, output and total tokens from the usage metadata. - Compare cost per successful task, not cost per call. Carbon's higher token use can still win if it removes retries.
- Route by difficulty: Argon by default, Carbon for the tasks where it changed the outcome.
Finish this testing before November 30 and the experiment itself runs at half the December price.
Settings that control the bill
thinking_level
Carbon accepts low, medium (the default) and high. Google recommends evaluating all three and picking the lowest level that meets your quality bar. Low uses the fewest tokens and is fine for lighter analysis and writing; high spends the most and is meant for the hardest math, science and code.
max_output_tokens
This one setting caps thinking plus answer together. The guide suggests at least 32,768 if your app currently sets something lower, because Carbon may spend much of a small budget on thinking and cut off the answer. The flip side: an unbounded long-decode run on a hard prompt can spend the whole allowance, so set a ceiling that matches what you are willing to pay per call.
Sampling parameters are ignored
The migration checklist says to remove temperature, top_p and top_k; if you send them, they are ignored. Changing them will not reduce cost or length.
Batch and Flex for anything that can wait
Nightly evaluations, document backlogs and report generation do not need instant answers. Both lanes cut the per-token price in half. Batch jobs have their own limits: 100 concurrent batch requests, a 2 GB input file limit, and up to 5,000,000 enqueued tokens for Carbon across active batch jobs.
Editor's note
Who it's for: teams already paying for the Gemini API who have hard, multi-step tasks (large refactors, deep document analysis, security review) where a failed attempt is expensive.
Not for: hobby projects on the free tier (Carbon and Argon are paid-only), or chat-style apps where speed matters more than depth; 3.8 Flash or Argon will be cheaper per useful answer.
Gotchas I found in the docs: the price doubles on December 1 for both Gemini 4 models; explicit caching code will not work on Carbon; thinking tokens are billed at the output rate.
Common error and fix: a project with a newly linked billing account sits in usage Tier 1, where a spend-based rate limit (if it applies to your account; Google says that depends on billing history) is of $10 per rolling 10 minutes, with a $250 billing tier cap. At Carbon's intro output price, $10 is about 1M output tokens, roughly one full long-decode job. Hitting it returns 429 RESOURCE_EXHAUSTED; the official fixes are to wait and retry, send shorter or smaller requests, or request a rate limit increase. Tier 2 (after $100 paid and 3 days from the first successful payment) raises the spend limit to $200 per 10 minutes.
FAQ
Is Gemini 4 Carbon free to try?
Not through the API: the pricing page lists it as not available on the Free tier, and the same goes for Argon. Google AI Ultra subscribers get it in the Gemini app, AI Mode in Search and Antigravity; researchers can apply to a separate access program described in Google's launch post.
What happens to my bill on December 1?
Standard rates go from $2 / $10 to $4 / $20 per million input / output tokens, and Batch, Flex, Priority and cached-input prices double too. Nothing in your code needs to change, but every estimate you made in November is off by a factor of two.
Is Carbon more expensive than Argon?
Not per token; the rate cards are identical. Per task it usually costs more, because Google says Carbon spends more thinking tokens per turn. Measure tokens on your own tasks before routing traffic.
Related on lightmind24: OpenRouter Pricing: Free Models, 5.5% Fee and BYOK · Groq API Free Tier: Rate Limits, Costs and When to Upgrade · Together AI Pricing: Credits, Rates and Limits
If you are going to benchmark Carbon against Argon, do it this month, while every test token costs half of what it will in December.
Screenshots: see the official pricing, model guide and rate limit page screenshots, plus my own cost-per-job table, at the end of this post. Images: Google AI for Developers documentation and lightmind24 own calculation.
Last verified: 2026-11-12
Sources: Gemini Developer API pricing · What's new in Gemini 4 Carbon · Gemini 4 Carbon model page · Gemini 4 Argon model page · Gemini API rate limits · Flex inference · Google blog: Gemini 4 Carbon is now available (Nov 10, 2026)
lightmind24 is an independent review site, not affiliated with Google.
댓글
댓글 쓰기