Gemini 4 Carbon doubles in price on December 1: the maths

Gemini 4 Carbon reached general availability on Google's paid API on November 10, and the figure everyone quoted was $2 for a million tokens of input and $10 for a million of output. Google's own pricing page carries the rest of the sentence: those rates hold only to the end of November. From December 1 they become $4 and $20.

So any pilot that runs this month and gets signed off on this month's numbers is wrong by a factor of two before it reaches production. The window is three weeks.

What it does to a bill

Take a modest production workload of 50 million input tokens and 10 million output tokens a month, none of it cached. At the introductory rate that is $100 of input and $100 of output, so $200 a month. From December 1 the same traffic costs $200 of input and $200 of output, so $400 a month. The step up is $200 a month, or $2,400 a year, on a workload that has not grown by a single token.

Now the part that catches people. Google's pricing table does not call that second line "output". It calls it output including thinking tokens, and the note underneath is blunt: every token the model processes gets counted and billed, reasoning included, and so do the internal tool instructions and function declarations that grounding adds. A reasoning model's invisible work lands on the invoice at the output rate.

Artificial Analysis, as reported by Lindy, measures Carbon writing about 123,000 output tokens per Intelligence Index task. At the introductory rate that single task costs $1.23 in output alone. From December 1 it is $2.46. Multiply by your task volume rather than by your word count.

The comparison nearly everyone got wrong

Carbon has been benchmarked all week against Claude and GPT. The comparison that decides a Google bill is Carbon against Gemini 4 Argon, and on Google's own pricing page the two are indistinguishable. Argon carries the same $2 and $10 to the end of November, the same $4 and $20 after it, the same $0.10 cached-input rate and the same rule about reasoning tokens being billed as output.

Identical per-token pricing means cost per job turns entirely on how many tokens each model burns. And Google's description of Argon is that it reaches the same Gemini 4 intelligence while answering faster and spending fewer tokens per turn than Carbon does, across coding, agentic, maths and science work.

Read that as a pricing statement, and note that it is Google's own efficiency claim about its own models rather than a third-party measurement. If it holds for your workload, then wherever Argon is good enough it is the cheaper model at the same list price, and the saving needs no negotiation. Carbon earns its place on the work that genuinely needs the stronger model: Google points it at complex software engineering, deep maths and science, high factual accuracy, autonomous multi-step agents and very long generations.

For what it is worth on the external comparison, Lindy reports Carbon and Claude Opus 5.5 both listing at $4 and $20 at standard pricing, with Artificial Analysis putting Carbon at $7.76 per task against $10.04 for Opus 5.5, about 23% less, because of how many tokens each writes. AtomicAgent gives the same $7.76 for Carbon and notes it already reflects the December rate, with Claude Fable 5.2 at $9.08. Treat all of those as third-party benchmark estimates rather than list prices.

Two levers, one of them blocked

The caching lever works, within limits. Cached input is $0.10 per million through November 30 and $0.20 after, against $2 and $4 for fresh input. Run the same 50 million input tokens with half served from cache and the November bill falls from $200 to $152.50, and the December bill from $400 to $305.

The limit is in Google's own footnote, which says Carbon supports "implicit caching only". Explicit caching is not available, so you cannot pin a long system prompt or a reference document into a cache and guarantee the discount. You can order prompts so the stable part comes first and hope the implicit cache catches it. Budget at the uncached rate and treat any cache hits as upside.

The second lever is the cheaper tiers, and it needs stating carefully, because every tier doubles. On AlphaCorp's reading of Google's table, Batch and Flex go from $1 and $5 today to $2 and $10 on December 1, and Priority from $3.60 and $18 to $7.20 and $36. Nothing is exempt from the rise.

What the tiers give you is a way to hold your current unit cost rather than avoid the increase. Batch stays at half of Standard, so asynchronous work moved off Standard before December lands on $2 and $10 in December, which is exactly what Standard costs today. Google's own page lists Batch, Flex and Priority alongside Standard for Carbon and describes its Batch API as halving cost, though only the Standard rates were readable in Google's documentation for this piece. For anything that does not need an answer in real time, that is the largest saving on the table, and it is a tier migration rather than a reprieve.

Two smaller items. Carbon is absent from the free tier, so there is no zero-cost path for experiments. And grounding bills separately: a monthly allowance of 5,000 free search requests, pooled across every Gemini 3 and later model rather than per model, then $14 for each additional 1,000, capped at 1,500 requests a day.

What IT teams should do

• Re-run every cost model at $4 and $20 before approving anything. A business case built on $2 and $10 expires on November 30.
• Measure billed output tokens, not visible output, on a representative sample of your own tasks. Thinking tokens are the gap between the estimate and the invoice.
• Run the same evaluation set through Argon and Carbon and compare cost per completed task. At identical list prices, the cheaper model is whichever finishes in fewer tokens.
• Identify every asynchronous workload and move it to Batch or Flex. Those tiers double too, but they land in December on the rate Standard charges today, which is the closest thing to holding your unit cost.
• Order prompts so the stable prefix comes first to give implicit caching a chance, but budget at the uncached rate.
• If you buy through Gemini Enterprise Agent Platform, get your own rates in writing. Google's page warns those prices may differ from the published API list.

Sources: Google's Gemini Developer API pricing, AI Pricing Guru, AtomicAgent, Lindy, FelloAI, AlphaCorp.