Gemini 4 Carbon went generally available in GitHub Copilot on November 10. For three weeks it's billed at an introductory rate: $2.00 per million input tokens, $0.10 per million cached input tokens, $10.00 per million output tokens. On December 1, every one of those numbers doubles to $4.00, $0.20 and $20.00.

That's the headline. The part worth sitting down with is what the doubling does to real work, because per-token prices are abstract until you multiply them by a task. Below are three task shapes with the arithmetic done out, using GitHub's current rate card and Google's December 1 rates. Every number is either from those two sources or plain multiplication on them, and I'll say which is which.

The rates

Token type Now (until Nov 30) From Dec 1
Input $2.00 / 1M $4.00 / 1M
Cached input $0.10 / 1M $0.20 / 1M
Output $10.00 / 1M $20.00 / 1M

One sourcing note, because it matters: GitHub's pricing page lists only the current rates. The December 1 rates come from Google. The launch post's footnote says the introductory period expires on November 30, 2026, after which $4/$20 apply, and Google's pricing page confirms the same split. Copilot bills at provider list pricing, so the rise flows straight into Copilot billing. I'm not going to write "GitHub lists the December rates" when GitHub's page doesn't.

Carbon is available on Copilot Pro+, Max, Business and Enterprise. Free and Pro don't get it. And code completions and next-edit suggestions aren't billed per token at all — they're unlimited on paid plans — so this whole price discussion applies to chat, agent mode and review, not to the completions firing as you type.

Three task shapes, priced

These three profiles come from a worked example published on dev.to on November 14. The token counts are that author's illustrative shapes, not measured usage, and the author is upfront about it: the figures are arithmetic on the unit prices and ignore extra thinking tokens, so treat them as a floor. I've recomputed each one and they check out.

A single chat prompt: $0.018 now, $0.036 from Dec 1

4,000 input tokens, 1,000 output tokens, nothing cached.

  • Now: 4,000 × $2.00/1M = $0.008, plus 1,000 × $10.00/1M = $0.010. Total $0.018.
  • From Dec 1: $0.016 + $0.020 = $0.036.

A thousand prompts: $18.00 now, $36.00 after. At this scale the doubling is real money only in aggregate. If Carbon is your quick-question model, December costs you an extra $18 per thousand prompts. Noticeable, not dramatic.

A typical agentic coding task: $0.444 now, $0.888 from Dec 1

60,000 fresh input tokens, 240,000 cached input tokens, 30,000 output tokens. This is the shape of an agent working through a real task with a big context window mostly served from cache.

  • Now: $0.12 + $0.024 + $0.30 = $0.444.
  • From Dec 1: $0.24 + $0.048 + $0.60 = $0.888.

A hundred such tasks: $44.40 now, $88.80 after. In Copilot's own currency, at 1 credit = $0.01, that's roughly 44 credits a task now, 89 after.

Look at where the money goes. Output is the smallest token count in this profile, 30,000 against 300,000 input, and it's two-thirds of the bill: $0.30 of $0.444. Cached input, the biggest count by far, costs almost nothing: $0.024. The lesson is that with Carbon, the expensive thing isn't the size of your context. It's how much the model writes back.

A long-horizon refactor: $2.60 now, $5.20 from Dec 1

800,000 input tokens, 100,000 output tokens, nothing cached.

  • Now: $1.60 + $1.00 = $2.60.
  • From Dec 1: $3.20 + $2.00 = $5.20.

This is the shape where the doubling bites. Ten refactors like this a week is $26 now and $52 after, about $100 a month extra for one developer doing serious agent-driven work. Nobody goes broke on it, but it's the difference between "cheap enough to not think about" and "worth thinking about."

The honest caveats

Three things the profiles don't capture, and all three push real costs up.

First, thinking tokens. Carbon reasons at length, that's the point of it, and thinking tokens bill as output. The profiles above count only the output you see. A task that looks like $0.444 on paper can cost meaningfully more once the model's internal reasoning is on the bill. Nobody has published a measured same-task comparison of Carbon against Argon, so I can't put a number on that gap. Don't let anyone sell you one.

Second, these are illustrative shapes, not measurements. The token counts were chosen by one developer to show the arithmetic. Your tasks will differ. The rates are real; the shapes are examples. If your agent runs use longer contexts or produce more output, scale the numbers accordingly. Output is the lever that moves the total fastest.

Third, the doubling is a list-price change, not a Copilot-specific announcement. GitHub hasn't published a Copilot-specific December 1 notice that I could find. The mechanism is that Copilot bills at provider list pricing, so Google's rise flows through. If Google extends the introductory period, the numbers change. Check the rate card before December, not after.

Carbon vs Argon: same price, different bill

Here's the comparison that surprised me most. On GitHub's rate card, Carbon and Argon cost exactly the same per token: $2.00/$0.10/$10.00 now, $4.00/$0.20/$20.00 from December 1. Argon's price doubles too. So any cost difference between the two models on the same task comes from how many tokens each one uses, not from the rate card.

Google's own guidance is blunt about which one to use where: "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 — complex, multi-step coding and debugging, open-ended research. Carbon uses more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops, in Google's words. More tokens at the same price per token means a bigger bill per task, even before December.

Google also claims Carbon tops CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than twice lower cost per task than the next best external model. Those are Google's claims, not independent verification. And the same leaderboard that puts Carbon first puts it at $9.22 per task against Argon's $4.57. Better results, roughly twice the cost per task, on that one benchmark's numbers. That's the trade in a sentence.

For context, the cheaper sibling on the same rate card is Gemini 4 Hydrogen: $1.50 input, $0.075 cached, $7.50 output, with no December 1 change stated. If your work doesn't need Carbon's depth, Hydrogen is the budget option and Argon is the middle.

What I'd actually do before December 1

Try Carbon now, on the hardest tasks you have, while the introductory rate lasts. That's the window where experimenting is cheapest, and it's the only way to learn whether its extra reasoning is worth its extra tokens on your work. Then budget at the doubled rate: if a task shape costs $0.888 in December and you'd still pay it, keep using Carbon for that shape. If you wouldn't, that shape belongs to Argon or Hydrogen.

And keep the completions separate in your head from the chat costs. Completions are unlimited on paid plans, so the December change doesn't touch the part of Copilot that fires while you type. It touches the deliberate, agentic work, the runs where you hand the model a task and let it run. That's where the money is, and that's where the doubling lands.

Sources

P.S. Everything I write stays free — that part doesn't change. But if this one was worth something to you, you can buy me a coffee ☕ (one-time or monthly, both work). It covers the API credits and hosting behind these experiments — and honestly, it's also the nudge that gets the next one started on the evenings I'd otherwise skip it. — Salam

Buy me a coffee

AheadKit on Gumroad — Beginner-Friendly Books & Packs