If someone built you an agent, a support bot or a document workflow on Gemini 4 Carbon in the last week, the price they quoted you has an expiration date. Google launched Carbon on Nov. 10, 2026, at an introductory API price of $2 per million input tokens and $10 per million output tokens (Google). That price runs through Nov. 30. Starting Dec. 1, it's $4 and $20 (Gemini API pricing).
Same code, same traffic, twice the token bill. If the prototype math was done at the intro rate, your December invoice won't match it.
What Google actually said
Google calls Carbon its strongest model, built for long, multi-step agent work, and says it's available through the API in Google AI Studio and Gemini for business "for all paying developers and enterprises" (Google). The launch post's footnote is the line to pin above your desk:
After the introductory period expires on November 30, 2026, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.
That's from Google's launch post, which adds that it "applies to the Carbon model intended for public access only." The cybersecurity version Google offers through its Fairwind Program is a separate track.
We checked the details that change the math:
- "Doubles" is exact. Every Carbon token price on the pricing page goes up by precisely 2x on Dec. 1: input $2 to $4, output $10 to $20, cached input $0.10 to $0.20 (Gemini API pricing). The Batch, Flex and Priority tabs double too. Grounding fees stay put.
- No prompt size tiers. Some older Gemini models charge more once a prompt passes 200k tokens. Gemini 3.1 Pro Preview, for example, goes from $2 to $4 input above that line (Gemini API pricing). Carbon's table has one flat rate per tab.
- Thinking counts as output. The output line reads "including thinking tokens," and Google says every token is billed, including thinking tokens (Gemini API pricing).
- The model string is stable. It's
gemini-4-carbon, listed as the stable version, with no preview suffix (Gemini API docs). - No free tier. Carbon is "not available on the Free tier" (Gemini API pricing).
- These are Gemini Developer API prices. Google's own note says prices on Gemini Enterprise Agent Platform, its enterprise platform, may differ (Gemini API pricing). If your build runs there, check that price list separately.
One more thing from Google itself. It says Carbon uses "more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops" (Google). Google also claims Carbon tops several outside benchmarks at "more than 2X lower cost per task" than the next best external model. That's a vendor claim about frontier rivals, not about your invoice. A thorough model is a wordy model, and wordy is billed by the token.
The worked example
Here's a hypothetical. A 12 person services firm runs a customer support bot on Carbon. Our assumptions:
- 20,000 requests a month
- 6,000 input tokens per request (system instructions, retrieved help articles, conversation history)
- 1,500 output tokens per request, thinking included
Monthly volume first:
- Input: 20,000 × 6,000 = 120,000,000 tokens, or 120 million
- Output: 20,000 × 1,500 = 30,000,000 tokens, or 30 million
At the intro price ($2 / $10) (Gemini API pricing):
- Input: 120 × $2 = $240
- Output: 30 × $10 = $300
- Total: $540 a month
At the Dec. 1 price ($4 / $20) (Gemini API pricing):
- Input: 120 × $4 = $480
- Output: 30 × $20 = $600
- Total: $1,080 a month
To run your own numbers: (requests × input tokens ÷ 1,000,000 × $4) + (requests × output tokens ÷ 1,000,000 × $20). Pull real token counts from your usage page, not from a guess.
What the base math leaves out:
- Search grounding. If the bot uses Grounding with Google Search, you get 5,000 free search requests a month, shared across Gemini 3 and newer models, then $14 per 1,000, with a cap of 1,500 requests a day (Gemini API pricing). At one search per request: (20,000 - 5,000) × $14 ÷ 1,000 = $210 a month on top. Google notes one request can trigger more than one search. That fee doesn't change on Dec. 1.
- Caching. Cached input drops to $0.20 per million after Dec. 1, but Carbon supports implicit caching only (Gemini API pricing). You don't control when it kicks in, so we leave it out of the estimate. Treat it as upside.
- Batch. The Batch tab is half price: $2 / $10 from Dec. 1 (Gemini API pricing). A live support bot can't wait on a batch job. An overnight invoice extraction workflow can, and at these volumes it would cost $540 instead of $1,080.
The cheaper move
Switching to Argon doesn't dodge anything. It carries the same $4 / $20 price from Dec. 1 (Gemini API pricing), though Google says it spends fewer tokens per turn than Carbon.
Gemini 4 Hydrogen is listed at a flat $1.50 input and $7.50 output, with no end date attached (Gemini API pricing). Same workload:
- Input: 120 × $1.50 = $180
- Output: 30 × $7.50 = $225
- Total: $405 a month, versus $1,080 on Carbon
For a support bot answering shipping and refund questions, test Hydrogen against your real transcripts before paying for the model Google pitches at "open-ended scientific research and discovery" (Google). Fewer tokens of thinking would also lower the output count, but we kept the assumptions identical so the comparison is clean.
A warning if your freelancer suggests Gemini 3.8 Flash. It's $0.75 / $3.75 through Dec. 31, then $1.50 / $7.50 from Jan. 1, 2027 (Gemini API pricing). That's a teaser price too, just with a later expiration date.
What to do this week
Find what calls Carbon. Search your code and your automation tools for the string gemini-4-carbon. In Google AI Studio, the Usage page groups requests by named API key in a project, and lumps unmatched ones under "(unknown keys)," including keys without a display name (Google AI Studio docs). Name your keys so the report means something. API keys inherit billing from their project (Gemini API billing).
Check the spend. Google's billing guide says to monitor usage in AI Studio under Dashboard > Usage. For cost, open your Cloud Billing account's Reports page, group by SKU and filter Services to Gemini API (Gemini API billing).
Set a budget alert. In the Google Cloud console, go to Billing, then Budgets & alerts, click Create budget and choose Alerts only. Default thresholds are 50%, 90% and 100% of the amount you set (Google Cloud). Google warns that alerts-only budgets don't cap spending. For a hard stop, AI Studio offers an experimental project spend cap on the Spend page under Monthly spend cap > Edit spend cap, with possible overages during about 10 minutes of billing lag (Gemini API billing).
Check your tier ceiling. A Tier 1 billing account caps Gemini API spend at $250 a month, and when the account hits its cap, service pauses for every linked project until the 1st (Gemini API billing). On a Tier 1 account, our hypothetical $1,080 bot would hit that cap in about the first week of December. Tier 2 raises the cap to $2,000 once you've paid $100 and waited 3 days, and Google takes requests for higher limits (Gemini API billing). On Prepay, a $0 balance stops every key at once (Gemini API billing). Our surprise AI bill guide covers caps across other tools, and our earlier Gemini post covers the consumer plans.
Send us your usage export and we'll rebuild the December number for your workload: team@lirium.ai.