Google's strongest AI model goes on sale at a price that doubles on December 1

Google put its strongest AI model, Gemini 4 Carbon, on general sale on November 10. Paying developers and businesses can now use it through Google's Gemini API and its business products. Consumers on the top-tier Google AI Ultra plan are getting it in the Gemini app and in AI Mode in Google Search.

The headline for buyers is the price, and how briefly it lasts. Until November 30, Carbon costs $2 per million input tokens and $10 per million output tokens. A token is a small chunk of text, roughly three-quarters of a word. On December 1 both rates double, to $4 and $20.

Even at the higher rate, Carbon is cheaper per word than Anthropic's flagship. VentureBeat lists Claude Fable 5.2 at $8 and $40, and OpenAI's GPT-6 Astra at $10 and $50. Carbon's December price matches Anthropic's Claude Opus 5.5.

But independent testing suggests the per-word discount shrinks once you measure finished work. Carbon also arrives no longer the top-ranked model it was when Google first showed it off.

A three-week discount, then a December repricing

Carbon was announced on October 28, but only vetted cybersecurity teams could use it at first. Google still keeps its hacking-related abilities behind an application-only program called Fairwind.

The introductory window gives companies about three weeks of cheap testing. Any pilot that runs into December will see its bill per token double partway through. Finance teams budgeting from November invoices should plan for that.

Google is also unusually frank about who should buy it. Its launch post says the cheaper Gemini 4 Argon "remains the ideal choice" for most software and knowledge work. Carbon, it says, uses more tokens than Argon to explore and double-check its answers. Google pitches it for the hardest jobs.

Cheaper per word, not always per job

Artificial Analysis, an independent testing firm, now ranks Carbon fourth of 74 models on its main intelligence ranking, with a score of 60. The three spots above it all belong to Anthropic's Fable 5.2, released November 9, at different effort settings. An effort setting controls how long a model thinks before answering, and so how much it costs.

That is a fall from the top. When Google unveiled Carbon in late October, it scored 59 and led Artificial Analysis's ranking of 52 models at the time, according to NeuralTrust's review of the results.

Artificial Analysis lists Carbon at its $4 and $20 standard rate. On that basis, it estimates Carbon at its highest setting costs about $7.76 per test task. Fable 5.2 at its maximum setting costs $26.54 per task and scores five points higher.

The comparison tightens at matched settings, according to VentureBeat's reading of the same data. Fable 5.2 at high effort costs about $9.08 per task and scores 62, two points above Carbon. At medium effort, Fable is the cheaper of the two, about $3.80 per task against $6.59 for Carbon, though it scores two points lower, 56 to 58.

The reason is volume. Carbon used about 205 million output tokens to complete the full Artificial Analysis test run. NeuralTrust puts the median model at about 64 million. A lower price per word buys less when the model writes three times as many words.

Zapier's office-automation test tells a similar story. Carbon at high effort completed 53.9% of 657 tasks, at about $2.12 each, priced at the $4 and $20 rate. The medium setting finished 53.0% for $1.65 a task, almost as good for much less. Zapier did not publish comparable cost figures for Anthropic's models.

On that test, Google's cheaper Argon, at its high setting, passed 51.3% of tasks for about $1.70 each. That narrow gap helps explain why Google steers most buyers away from its own flagship.

What buyers still can't check

Several of Google's strongest claims lack published numbers. VentureBeat found no Carbon results in the public table for one coding test, CursorBench, where Google claims first place. Google's own results on hacking-related tests have not been reproduced by outsiders.

Rivals still lead on important work. On the FrontierSWE software-engineering test, Fable 5.2 scored 69.9% against Carbon's 61.2%, per VentureBeat. Fable also led on a separate coding test and on science tasks.

There is a reliability question too. Artificial Analysis measured Carbon giving confidently wrong answers 11% of the time on its fact-recall test, against 4% for Argon, according to NeuralTrust's summary. Carbon answered more questions correctly, but guessed more often when it did not know.

Practical gaps remain. Google has not published which countries can use Carbon, or whether its Vertex AI cloud platform carries it. Unlike open models such as Moonshot's Kimi K3.1 Pro, Carbon cannot be run on a company's own servers.

For buyers, the takeaway is to price Carbon per finished job, not per word, and at December rates. Google's own advice is that most of that work belongs on Argon. The model is a credible rival to Anthropic's best at a lower sticker price, but not a clear win on cost per task.