Gemini 4 Carbon gets twice as expensive on 1 December, and so does its cheaper sibling. Google opened Gemini 4 Carbon, its strongest model, to every paying Gemini API developer on 10 November at an introductory price of $2 per million input tokens and $10 per million output tokens. That rate runs through 30 November. From 1 December the standard price is $4 and $20, and Google's pricing page lists exactly the same change for Gemini 4 Argon.
For a solo developer or a small team, that leaves under two weeks to work out which jobs genuinely need Carbon, which can run on Argon, and how much of the bill can move to the cheaper ways Google sells the same models.
What Changes on 1 December
Google announced the wider release in a post by Koray Kavukcuoglu, the SVP of Google DeepMind and Google's Chief AI Architect. Carbon launched at the same introductory price as Gemini 4 Argon, and a footnote on the post set the end date: once the introductory period expires on 30 November, $4 per million input tokens and $20 per million output tokens apply.
The Gemini API pricing page spells out the full schedule. On the standard tier, input moves from $2 to $4 per million tokens and output, which includes thinking tokens, from $10 to $20. Cached input rises from $0.10 to $0.20 per million tokens. The Argon entry on the same page carries the same numbers on the same dates, so moving everything from Carbon to Argon does not avoid the increase. What it changes is how many tokens each request uses.
Neither model has a free tier. Both are listed as not available on the Free tier, so every call needs a paid Gemini API key, a Gemini for business plan or, for Carbon inside Google's own apps, a Google AI Ultra subscription.
The Same Price Per Token Is Not the Same Bill
Google is open about the difference between its two models. In the announcement, Kavukcuoglu wrote that "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". Carbon gets there by "using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops".
Those extra tokens are billed. The pricing page says every token the model processes is counted, thinking tokens included, and thinking is charged at the output rate. Google's developer notes describe the two models as sharing the same context window, built-in tools and per-token pricing, with the main difference being how many thinking tokens each one spends per turn.
Real Python measured what that looks like on a tiny request. Asked to explain a Python generator in two sentences, Carbon wrote a 69-token answer after spending 448 tokens thinking, at its default medium setting. Thinking made up 448 of the 517 tokens billed at the output rate, about 87 percent. At the introductory price the call cost about half a cent, which is nothing on its own and a real sum across millions of requests. The tutorial sums the model up in one line: "Carbon is Argon with more thinking: same context window, same tools, same price per token."
DataNorth, a Dutch AI consultancy, drew the practical conclusion in its briefing on the release: "The same price per token therefore does not mean the same bill." A task that Argon finishes in one pass may cost noticeably more on Carbon, even though both are billed at identical rates.
Who Can Use Gemini 4 Carbon Today
Carbon reached a small group first. Two weeks before the wider release, Google gave it to a set of cybersecurity defenders through its Fairwind Program, and the specialised cyber version stays on that separate route. The general model is now open on the paid tier of the Gemini API in Google AI Studio and Android Studio, and to companies through Gemini for business, Google Antigravity and CodeMender.
Consumers get it through Google AI Ultra, in the Gemini app, AI Mode in Search and Antigravity. DataNorth pointed out that the announcement does not mention the cheaper AI Pro plan. Scientists have a separate door: four weeks before the launch, Google opened applications to give 10,000 scientists free access to Argon in Antigravity for six months, and it has now extended that programme to Carbon.
There are limits on what access means. Superpower Daily noted that Carbon is a hosted service with no downloadable weights or self-hosting licence announced and, citing VentureBeat, that Vertex AI availability, regions and support commitments were left unspecified. DataNorth made a similar point for European buyers: Google does not say where Carbon processes data for Gemini for business customers.
Four Ways to Keep the Bill Down After 1 December
The increase is real, but Google's own price list leaves several levers for teams that watch every dollar.
Use Batch or Flex for Work That Can Wait
Batch and Flex requests cost half the standard rate. Today that is $1 per million input tokens and $5 per million output tokens; from 1 December it becomes $2 and $10. In other words, from December the Batch and Flex tiers cost exactly what the standard tier costs now. Overnight evaluations, document processing and anything else that does not need an answer within seconds belongs there. At the other end of the list, the Priority tier rises from $3.60 to $7.20 per million input tokens and from $18 to $36 per million output tokens.
Turn the Thinking Down
Carbon offers three thinking levels, low, medium and high, all at the same per-token price, with medium as the default. In Real Python's test, Carbon answered a simple letter-counting question correctly at both ends of the scale, skipping thinking entirely at low and spending 272 thinking tokens at high. Google's advice, as Real Python reported it, is to try a workload at each level and keep the lowest one that is good enough.
Send Routine Work to Argon
Google's developer notes describe Argon as faster and lighter on token spend per turn than Carbon for interactive coding, day-to-day feature work and fast production loops, and reserve Carbon for complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing. Both models double on the same day, so the saving comes from Argon's lighter thinking, not from a lower rate.
Let the Cache Work and Watch the Output Cap
Cached input costs a twentieth of the uncached rate: $0.20 per million tokens from December, against $4. Caching is implicit for both models, and explicit caching is not supported. One trap catches people who budget tightly: the output cap covers thinking and the answer together. Real Python set a 256-token limit, and Carbon spent 253 tokens thinking, then returned an empty answer marked incomplete without raising an error.
Test Now, Budget at December Prices
The introductory window is most useful as a cheap trial. DataNorth suggests taking ten to twenty real tasks, running each on Argon and on Carbon, and logging the success rate and the tokens used per task. "Then price the results at $4 and $20, not at today’s rate," the consultancy wrote. If Carbon only wins a handful of those tasks, its advice is to send that handful to Carbon and keep the rest on Argon.
Google's benchmark claims are worth testing rather than taking on trust. The announcement says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next best external model. DataNorth observed that the post gives no scores, and that VentureBeat found no Carbon row in the public CursorBench table, while Zapier's own numbers back the AutomationBench claim. Carbon's result on Google's 375-task honesty test, 92 percent of tasks passed against 70 percent for Argon, comes from Google's own benchmark, and DataNorth noted that no outside party has checked it yet.
What the Increase Means for Small Builders
Large companies can absorb a price rise in a budget line. A freelancer running an agent on a client project, or a two-person start-up whose product calls the API thousands of times a day, feels it at once. Google published the end date on launch day, so the rise is not a surprise, and the levers it sells alongside the higher price are real: half-price Batch and Flex, adjustable thinking, automatic caching and a lighter model for routine work.
At MW3.biz, we believe the most capable tools should be within reach of a two-person team as well as a large enterprise. Clear, published pricing helps with that, because small builders can plan around a number they can see. From a technology-democratisation perspective, the healthiest outcome is a market where frontier models keep a lower-cost route, like the Batch tier, for people who build on tight budgets. That is one lens among many; teams with bigger budgets may simply see the introductory period as the discount it always was.
The same question, what a model really costs once list prices apply, shaped our look at Cursor and Claude Code at API prices. For Gemini 4 Carbon, the practical answer is straightforward: measure your real tasks before 30 November, price them at $4 and $20, and move the work that does not need deep reasoning to Argon or to the Batch tier before December arrives.