Gemini 4 Carbon cost per task: Google's 2X claim, checked board by board at the December price

Gemini 4 Carbon became generally available on November 10, and Google’s announcement puts cost per task ahead of the price:

It ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external model.

Cost per task is the right unit for agent work, because two models at the same rate per token can spend very different amounts finishing the same job. But a cost per task is tokens times a price, and Carbon has two prices: an introductory one until November 30 and a doubled one from December 1. Every component doubles, input, cached input and output alike, so a cost per task measured at the introductory rate is half of what the same run costs in December.

So the question for each leaderboard is which price it applied. The answer differs by board, and it moves Google’s claim. I have not run Carbon; everything below is relayed from the linked pages, and the ratios and December figures are my arithmetic on them.

What does Gemini 4 Carbon cost per token, before and after November 30?

The same as Gemini 4 Argon: $2 per million input tokens and $10 per million output through November 30, 2026, then $4 and $20. Cached input is $0.10, then $0.20, and only implicit caching is supported. Thinking is billed as output. Batch and Flex are half the standard rate. Carbon is not on the free tier, and the pricing page has no long-prompt surcharge for it.

What differs from Argon is how much Carbon writes. Google’s migration page says Carbon spends more thinking tokens per turn, and the announcement says Argon stays the better choice for most software engineering and knowledge work. Carbon, in its words, is for harder tasks:

using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops.

More tokens per turn at the same rate. Google’s claim is that this still comes out cheaper per task than the competition.

Which Gemini 4 Carbon price did each leaderboard use?

Two used the introductory rate, two used list, one does not say. Here are the five boards that publish a Carbon cost per task, with the December figure where the basis allows one. All Carbon rows are the High setting.

Google cites the first two for its 2X (and WANDR, which has no public Carbon cost, below), and MKA for its 6X. Agent Arena and the Intelligence Index are not part of Google’s claim; they are here because they publish the same unit. The December column assumes the run would use the same tokens, and that the competitor’s price does not change.

What does CursorBench 4.0 show for Gemini 4 Carbon?

A narrow first place, priced at the introductory rate. CursorBench runs agents on ambiguous multi-file tasks taken from real Cursor sessions. Its leaderboard has Carbon High at 60.4%, Carbon Medium at 60.1% and Claude Fable 5.2 Max at 60.0%, and warns that small differences may not be statistically meaningful.

The page computes cost by applying “published” per-token pricing and links to Cursor’s models and pricing page. That page lists Carbon at $2 input, $0.10 cache read and $10 output, with the December rates in a note. If those are the rates behind the leaderboard, Carbon High costs about $18.44 per task from December 1, and the multiple against Fable 5.2 Max falls from 3.6 to about 1.8. The leaderboard does not record the date of each run, so this is the rate card’s reading, not a confirmed one.

Two more rows matter if you are choosing a setting. Carbon Low costs more per task than Medium, $11.89 against $8.49. Per step it is cheaper, about 10 cents against 13, but it took nearly twice the steps. And Opus 5.5 High scores 56.0% at $3.97, four points below Carbon Medium for less than half its introductory cost per task.

Carbon against Argon is the row I cannot explain. Both are on the same rate card, and Carbon reports fewer tokens per task (37,067 at Medium against Argon’s 103,600) and fewer steps, yet costs about twice as much. The page does not split cost into input, cache and output, so the reason is not visible from it.

What does Zapier’s AutomationBench show for Gemini 4 Carbon?

The one clean confirmation: first place, priced at list, 2.2 times cheaper than the next model. AutomationBench runs end-to-end business workflows across 47 simulated apps and grades the final state of the environment, strict pass or fail, with no model as judge.

The leaderboard footnote covers only the Argon rows, but Zapier’s separate Carbon report states the basis: $4 input and $20 output per million tokens, and cache reads assumed at $0.40. Google lists cached input at $0.20 from December, so if anything this overstates Carbon’s cost. The 2.2 holds in December.

The report adds a figure the leaderboard does not show: spend per pass, all task costs divided by the tasks passed, failures included. For Carbon High it is $3.94, against $2.12 per task. The run took place on October 26 and 27, two weeks before general availability. Note too that on this board Opus 5.5 and GPT-6 Astra cost less per task than Carbon High, and score eleven to twelve points lower.

What do Perplexity’s WANDR and the MKA benchmark show for Gemini 4 Carbon?

WANDR has no public Carbon cost, and MKA shows 5X where Google says 6X. The WANDR post Google links is dated July 14, 2026, four months before Carbon’s release, and has no Carbon row. The #1 ranking and the cost multiple on WANDR rest on Google’s word.

MKA was built by Google DeepMind and is run by Artificial Analysis. Their launch article of November 10 has Fable 5.2 Max first at 57% and averaging $31.97 per task, and Carbon High second at 54% and $6.29. That is 5.1 times lower. Google’s announcement says “more than 6X lower” against the same top model. The article does not say which Carbon price it used, so I cannot tell whether the gap comes from the price basis or from something else.

What does Agent Arena’s cost per task for Gemini 4 Carbon measure?

Real user tasks at the introductory price, and only the start of each session. Agent Arena ranks agents on tasks people bring to it. Its overall leaderboard, dated November 8, has Carbon second and marked preliminary:

At $2 and $10 Carbon’s median task costs about a quarter of Opus 5.5’s. At $4 and $20 it would be about $2.92, roughly half. Opus still ranks first. Against Argon, Carbon writes nearly twice the output per task and costs nearly twice as much, which fits Google’s “more tokens than Argon” better than the CursorBench rows do.

What a “task” covers matters here. Arena’s cost post from August explains the cap it applies:

To measure costs in a representative manner, we decided to track costs for, at most, the first three tasks in an Agent Arena session.

Arena’s reason is context carry-over: later tasks in a long session cost more whatever the model, and it did not want that to punish models people keep talking to. The consequence is that the cost columns leave out the part of a long session where context piles up. Arena’s methodology post has a general finding, without naming models:

We find some models more expensive in practice, despite cheaper on-paper pricing.

How does Gemini 4 Carbon compare with Argon on the Artificial Analysis Intelligence Index?

At the same cost per task. Artificial Analysis reported its Carbon evaluation on October 28, before the public release, and put Carbon at high reasoning first on its Intelligence Index with 59. At standard pricing, which it says all its charts use, Carbon cost $7.76 per index task, Argon $7.68, Claude Opus 5.5 Max $10.04 and GPT-6 Astra Max $6.25. Google’s 2X is not claimed for this index. Here Carbon is about 23% cheaper per task than Opus 5.5 Max and costs more than GPT-6 Astra Max.

What does Google say a Gemini 4 Carbon /swarm run costs?

Nothing in numbers, only a warning about quota. The Antigravity post describes /swarm, in preview for Google AI Ultra in the Antigravity app and CLI: a coordinator starts parallel workers in isolated git worktrees, adversarial verifiers run tests and challenge diffs, and all of them post to a shared message board, idle at zero tokens until needed. Carbon takes the coordinator, architecture and auditing seats; lighter models do routine edits. Google reports five weeks of internal use on migrations and root-cause investigations, with no metrics, and adds:

(As a fair warning, even with zero-token idle waits and dynamic model hot-swapping, parallel multi-agent systems consume quota much faster than a standard chat.)

That is subscription quota, not an API bill, and it is the closest thing to a long-run cost Google publishes. Simon Willison’s launch notes have no token counts or run costs either.

What do the leaderboards leave out about a long Gemini 4 Carbon agent run?

The run itself. Every number above is an average or a median per benchmark task: about 71 steps on CursorBench for Carbon High, one business workflow on AutomationBench, at most the first three tasks of a session on Arena. None of them reports what one multi-hour Carbon session costs, or how Carbon’s verification loops grow with the length of the job.

They also disagree on Carbon against Argon, which is the comparison Google itself invites by recommending Argon for most work. Per task Carbon costs about twice Argon on CursorBench and Arena, 1.25 times on AutomationBench, and the same on the Intelligence Index. The rate is identical on every board, so the spread comes from the tasks, and none of those tasks is yours.

Should you move a long agent run to Gemini 4 Carbon before November 30?

Test it before November 30, decide at the December price. This is a reading of the sources, not a measurement.

An introductory price with an end date is a discount on testing, not on the workload, and the same rule applies to the leaderboards. Two of the five already price Carbon at the standard rate. Double the CursorBench and Arena figures before comparing them with anything priced at list, and leave the MKA multiple aside until its basis is known.

Then measure your own denominator. Log cost per finished task on your own runs, failures included: Zapier’s own report shows the gap between $2.12 per task and $3.94 per pass. Run it at all three thinking levels, since CursorBench’s Low row is the most expensive per task. The Haiku 5.5 subagent post found the same pattern from the cheap end: the same price per token as GPT-6 Luna, and about three times the output tokens per task at max effort.

None of the Carbon sources publishes a review record either. For Argon, the one Google case that did kept two of six proposals, which is why I argued that review capacity is the first thing to size, before tokens. For Carbon, the share your reviewer keeps is yours to measure.