Wain AI/Tech Blog

AI news and trends worldwide, updated nearly every day

Gemini 4 Carbon Reaches General Availability - Introductory Price Ends November 30, $4 Input and $20 Output From December 1

Image: Google official blog

Gemini 4 Carbon Reaches General Availability - Introductory Price Ends November 30, $4 Input and $20 Output From December 1

On November 10, 2026, Google opened Gemini 4 Carbon to developers, enterprises, and consumers. The paid-tier API costs $2 input and $10 output through November 30, then $4 and $20. Rates match Argon, but Google says Carbon spends more tokens. A /swarm preview began in Antigravity.

On November 10, 2026, Google widened access to Gemini 4 Carbon (model ID gemini-4-carbon), making it open to consumers, enterprises, developers, and scientists 1. When Google first announced the model on October 28, only a subset of cyber defenders could get it, and the end of the introductory period had not been disclosed. Both of those have now changed. Per million tokens, the introductory rate of $2 for input and $10 for output lasts through November 30, 2026, and from December 1 the rate is $4 and $20 2. That leaves about three weeks of introductory pricing, counted from November 10.

The Unit Price Doubles Between November 30 and December 1

The Gemini API price list shows two rates for each processing class, one through November 30 and one from December 1 (USD per million tokens) 2.

ClassInputOutput (including thinking tokens)Cache
Standard (through 11/30)$2.00$10.00$0.10
Standard (from 12/1)$4.00$20.00$0.20
Batch and Flex (through 11/30)$1.00$5.00$0.05
Batch and Flex (from 12/1)$2.00$10.00$0.10
Priority (through 11/30)$3.60$18.00$0.18
Priority (from 12/1)$7.20$36.00$0.36

Every class in this table doubles on December 1. Batch and Flex stay at $2 input and $10 output after that date, the same amount Standard charges during the introductory period. The Free Tier column reads “Not available,” so API use assumes the paid tier. Caching is implicit only; explicit caching is unsupported.

Carbon’s switchover date matches the schedule set out when Argon became generally available. On the same price list, Argon’s Standard rate also goes from the same $2 and $10 to $4 and $20 on December 1. A setup that mixes the two models will see both unit prices double at the start of the month.

A footnote in the announcement says this pricing applies only to the public Carbon model 1. The announcement treats Carbon for cyber separately, offering it on a rolling basis to those who apply to the Fairwind Program, and neither the announcement nor the price list states what that version costs. Both documents also give only dates, with no hour or time zone for the switch. If a large batch job will run across the end of the month, it seems safer not to assume the introductory rate holds until late on November 30.

Eligibility Depends on the Route

The API is offered through AI Studio and Gemini for business to every paying developer and enterprise, according to Google 1. Google’s developer guidance also names Android Studio as a place to use the paid-tier Gemini API. Among individual subscriptions, the eligible plan is Google AI Ultra, and the rollout to Antigravity, the Gemini app, and AI Mode in Search has begun. AI Ultra subscribers can also try the model in AI Studio.

This consumer scope is where Carbon differs from Argon. Google’s October 13 Argon announcement included AI Pro subscribers for the Gemini app and AI Mode. Carbon’s announcement names only AI Ultra for consumers and says nothing about AI Pro.

Enterprises can reach Carbon through the API or the model menu in Gemini for business, Antigravity, and CodeMender. For researchers, Google runs an application process that may grant six months of free use of Argon and Carbon alike on Antigravity’s top usage tier, with eligibility rules, limited capacity, and a selection step.

Same Per-Token Rate, More Tokens Than Argon

Google positions Carbon as its strongest model, while also saying Argon is still the better fit for most engineering and knowledge tasks 1. The uses it lists for Carbon are intricate coding and debugging across many steps, along with exploratory science, and it says the model consumes more tokens than Argon as it explores and verifies. The price list likewise describes Argon as quicker to respond and lighter on tokens per turn than Carbon, and it bills thinking tokens at the output rate 2. Matching rates per million tokens do not mean matching bills for the same job.

Google also puts numbers on its cost-per-task case. According to the announcement, Carbon placed first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, with a per-task cost under half that of the runner-up external model 1. On the Multimodal Knowledge Agent benchmark that Google launched with Artificial Analysis, Carbon placed second, with an average per-task cost under one-sixth that of the top model. All of these are figures from Google’s announcement, and the announcement text does not say which unit price went into the cost calculation. If the introductory price was used, the gap in the same comparison would narrow after December 1, as long as the rivals’ prices stay put.

Security use calls for a separate caution. Google attaches a condition to Carbon’s lead on SEC-bench Pro, ExploitGym, and CWE-bench v1: access via the Fairwind Program 1. The announcement does not present those figures as ones that carry over directly to the public Carbon on the API.

Antigravity’s /swarm Is a Preview That Uses Quota Quickly

Alongside the wider release, a preview of /swarm began in Antigravity. According to the documentation, running /swarm turns the current conversation into a coordinator, which divides the job and hands the pieces to named workers 3. All members talk on a shared board, and a @verifier is usually added for each group of two to four workers to look over diffs and run tests. A comparison table on the same page recommends /swarm for jobs that split cleanly by file, package, or question, and /teamwork for migrating whole repositories, simulations that run for a long time, and research efforts. Our September comparison also covered how Antigravity runs multiple agents.

The plan conditions take some cross-reading. Both Google AI Pro and AI Ultra include /swarm itself in the Antigravity App and CLI 3. But workers default to the conversation’s model, and any other model requested has to be selectable in the model picker. Since the only individual plan Carbon’s announcement names for Antigravity is AI Ultra 1, it appears that an AI Pro /swarm cannot run Carbon workers.

On cost, the documentation itself warns that each worker is a separate agent, so quota drains much faster than in one conversation 3. Idle workers consume no tokens, but staffing an entire team with Carbon, which uses more tokens than Argon, would likely speed that drain further. The arrangement Google describes puts Carbon on roles requiring deep reasoning, such as design, coordination, and auditing, and mixes in quicker, lighter models 1. The announcement shows 3D exhibits and games built with /swarm, but these are Google’s own demonstrations, and it reports neither token counts nor costs for them.

What Three Weeks of Introductory Pricing Can Settle

Google has now given the release date that was missing in October. Even so, the window at the introductory rate is short, and the unit price to use for production estimates is the December 1 rate of $4 input and $20 output.

One practical use of this period is to send identical tasks through both Argon and Carbon and log the token counts, thinking included. With those counts in hand, a team can recalculate its bill at the December rates and limit Carbon to work that justifies the extra tokens. It is also a good moment to test whether non-urgent Standard jobs can move to Batch or Flex, which would halve the unit price after December 1.

Sources

  1. Gemini 4 Carbon is now available - Official Google blog (November 10, 2026)
  2. Gemini Developer API pricing - Google AI Studio (price list, checked November 11, 2026)
  3. Swarm (/swarm) - Official Google Antigravity documentation (checked November 11, 2026)

We publish the latest AI news nearly every day.

Subscribe via RSS Get new posts the moment they go live.

Search other keywords →