Google Launches Gemini 4 Carbon With 1M-Token Decodes | Zeniteq
Loading
Loading article
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Google Launches Gemini 4 Carbon With 1M-Token Decodes
Carbon’s November 10 broad launch brings paid API access, extended generation and temporary Argon-level pricing, with several migration details developers need to check.
Google made Gemini 4 Carbon broadly available on November 10, 2026, opening the model to paying developers and enterprises after a limited cybersecurity release through the Fairwind Program two weeks earlier.
Carbon's defining feature is a shared 1,048,576-token budget covering both input and output. Introductory API pricing matches Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, but those rates double after November 30. Several settings — including temperature, top_p, and thinking_budget — are either ignored or rejected, and the Antigravity /swarm feature remains a preview.
Google made Gemini 4 Carbon broadly available on November 10, 2026, opening its latest frontier model to paying developers and enterprises after a limited cybersecurity release two weeks earlier through the Fairwind Program. Carbon is also rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search, according to Google’s launch announcement.
The production model ID is gemini-4-carbon. Its defining technical feature is a shared 1,048,576-token budget for input and output, with long-decode support that lets generation continue across multiple segments.
Google is initially charging the same API rates as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens through November 30. The scheduled rates afterward are twice as high. That makes the launch window useful for evaluation, but a poor baseline for estimating ongoing production costs.
What Is Generally Available Now
Google lists Carbon as generally available through the Gemini API paid tier, accessible through Google AI Studio, and Gemini for business. The Google Cloud model documentation confirms the GA status and November 10 release date for gemini-4-carbon.
The documented feature set includes function calling, structured output, code execution, URL context, computer use and batch inference. Standard and priority pay-as-you-go access are supported, alongside provisioned throughput. Carbon accepts text, image, audio and video inputs, but its output modality is text.
The Ultra rollout is a separate availability track. Google says it is rolling out to subscribers, without specifying when every account will receive access. Developers can therefore distinguish an available paid API model from a consumer rollout that may still be reaching accounts.
Two other access distinctions matter: Antigravity’s /swarm feature remains a preview, while the specialized Carbon cybersecurity variant remains allowlisted. Neither should be treated as unrestricted access simply because the main Carbon model is GA.
The Million-Token Limit Is Shared
Carbon’s headline number needs careful reading. The model has one unified token allowance covering the prompt and generated output, including thinking, function calls and the final answer. It is not a million-token input window plus another million tokens of generation.
Google documents the following limits across its supported API paths:
Carbon limit
Documented allowance
Combined input and output
1,048,576 tokens
One foreground segment, streaming
Up to 65,536 output tokens
One foreground segment, non-streaming
Up to 32,768 output tokens
Extended output using long decodes
Up to 1,048,576 tokens minus input
These figures come from the Carbon model documentation. The smaller foreground limits explain why a nominally million-token model still needs continuation handling.
For example, an input consuming 262,144 tokens leaves a theoretical maximum of 786,432 output tokens. That remainder must accommodate thinking and function calls as well as the answer users see. A document-heavy task can therefore leave substantially less room for generation than the headline capacity suggests.
This is a useful distinction for repository migrations, research synthesis and large artifact generation. Adding more source material can improve the information available to the model, but it also reduces the remaining generation allowance. My view is that applications should reserve output capacity explicitly instead of filling the context window first and hoping the answer fits.
Long Decodes Require Continuation Handling
Google supports long decodes through the Interactions API and GenerateContent API, with an important version caveat: its documentation says continuation_token is supported only in v1beta1. The Google Gen AI SDK defaults to that version.
The documentation describes three ways to handle extended generation:
Google Gen AI SDK and Agent Studio: Automatically handle continuation tokens across segments.
Interactions API: Supports background=true for background generation.
Custom continuation loops: Pass the returned continuation_token to continue generation.
A custom client that stops after its first foreground segment will not receive the full long decode. Likewise, an application pinned to another API version needs to check continuation support before relying on it.
I’d make continuation behavior an early migration test. A short-answer smoke test can succeed while leaving the application’s long-output path completely untested.
The larger output allowance is also not exclusive to Carbon. Google’s migration comparison lists a 1,048,576-token output limit for both Gemini 4 Argon and Carbon, versus 65,536 for models such as Gemini 3.8 Flash. Carbon’s launch should not be read as the first appearance of that capability across the entire Gemini family.
Settings That Change During Migration
Switching the model ID is only part of the work. Carbon changes several controls that existing Gemini applications may assume are configurable.
Thinking levels:LOW, MEDIUM and HIGH are supported. The public gemini-4-carbon model defaults to MEDIUM.
Unsupported thinking settings: Supplying thinking_budget or setting thinking_level="MINIMAL" returns an HTTP 400 validation error.
Sampling controls: Custom temperature, top_p and top_k values are ignored. Their fixed values are 1.0, 0.95 and 64, respectively.
Caching: Carbon supports implicit caching with a 4,096-token minimum and persistent caching enabled by default. Explicit context caching is unsupported.
The difference between rejected and ignored settings is consequential. An unsupported thinking configuration produces a visible failure. A custom temperature value can remain in a request without controlling sampling as the application expects.
For teams using a low temperature as part of a consistency strategy, successful requests alone are insufficient evidence that the old behavior carries over. Evaluate the resulting outputs and use supported mechanisms, such as structured output, where appropriate.
Caching deserves a separate review. A pipeline built around explicitly created cache resources cannot assume those resources transfer to Carbon. Implicit caching offers reuse without explicit cache management, but the documented minimum is an eligibility condition, not a promise that every repeated request receives a particular saving.
Carbon also lacks Gemini Live API support and fine-tuning. Those are feature-selection constraints, not settings that can be restored by changing a request parameter.
Introductory Pricing Ends November 30
Google’s announcement sets out two pricing periods for the public-access model:
Pricing period
Per million input tokens
Per million output tokens
Introductory, through November 30, 2026
$2
$10
Scheduled after November 30
$4
$20
The introductory rate matches Argon’s announced pricing. It does not mean a Carbon task will cost the same as an Argon task: Google says Carbon uses more tokens to explore problems, reason and perform extended verification.
Consider an illustrative workload consuming 200,000 input tokens and 800,000 output tokens. Applying the published rates gives:
Introductory token charge: $0.40 for input plus $8 for output, totaling $8.40.
Scheduled later token charge: $0.80 for input plus $16 for output, totaling $16.80.
This is arithmetic using the listed token rates, not a measured Carbon workload or an all-in service quote. It excludes any applicable additional charges or discounts. It shows why output volume matters so much for long decodes: most of the bill in this example comes from generation.
The sensible evaluation question is therefore cost per acceptable result. If Carbon completes a difficult investigation with fewer failed attempts, a higher token bill might be worthwhile. If both models solve the task reliably, extra reasoning and verification may add expense without enough benefit. Those outcomes need workload-specific evaluation.
Google’s migration guide adds another accounting wrinkle: reported token counts may rise because upgraded infrastructure captures previously undercounted request components, including response schemas and function-calling metadata.
A before-and-after cost comparison should separate three changes: the model’s token consumption, more complete token reporting and the scheduled price increase. Otherwise, developers risk attributing the entire difference to Carbon’s reasoning behavior.
I’d budget against the post-November rates from the start. The introductory period reduces evaluation costs; the announcement provides no basis for assuming it locks in a lower ongoing rate.
/swarm Is a Preview; Carbon for Cyber Is Gated
Google is previewing /swarm in Antigravity as a way to turn one prompt into coordinated agent work. Its announcement describes assigning Carbon to demanding roles such as architectural design, coordination and auditing, while lighter models work in parallel isolated workspaces and adversarial verifiers check changes.
That division of labor is worth evaluating. The strongest model can be reserved for decisions where deeper reasoning matters, while other work uses lighter models. However, /swarm remains a preview; Carbon’s GA status does not extend automatically to the orchestration feature.
The cybersecurity variant has a different restriction. Google lists gemini-4-carbon-cyber-permissive as GA on its enterprise platform, but access is allowlisted and requires a separate application through Google or Fairwind. Previous access to other Gemini cyber models does not carry over.
The documentation also records operational differences:
Feature
Public Carbon
Carbon for Cyber
Access
Publicly available to eligible paying customers
Allowlisted
Default thinking level
MEDIUM
HIGH
Computer use
Supported
Unsupported
Listed locations
global, us, eu
global, us
The cyber variant is intended for defensive work including vulnerability discovery, reverse engineering, exploit validation and patch generation. Its GA label describes release maturity, not unrestricted availability.
Where Carbon Fits Beside Argon
Google positions Carbon for complex coding and debugging, long-horizon agent workflows and open-ended scientific research. It still recommends Argon for most software engineering and knowledge work.
The company reports first-place results on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, with more than twofold lower cost per task than the next-best external model in those comparisons. It also reports second place on the Multimodal Knowledge Agent benchmark, developed with Artificial Analysis, at more than sixfold lower average task cost than the top performer.
These are Google-reported benchmark and economic claims, not measurements reproduced for this article. The named evaluations provide context, but they do not settle performance or cost for a particular production workload.
Carbon’s most useful evaluation set will include tasks that expose its intended advantages: cross-subsystem debugging, multi-package migrations, dense-document analysis and research requiring repeated verification. For routine work, Google’s own positioning gives developers a reason to keep Argon in the comparison.
Frequently Asked Questions
4 questions
1
What Is the Gemini 4 Carbon Model ID?
The generally available public model ID is gemini-4-carbon. The allowlisted cybersecurity variant uses gemini-4-carbon-cyber-permissive.
2
Can Carbon Generate a Million Tokens in One Response?
Long decodes can extend generation to the shared 1,048,576-token limit minus input. Individual foreground segments remain capped at 65,536 output tokens when streaming or 32,768 when non-streaming, so continuation handling matters.
3
When Does Carbon’s Introductory Pricing End?
The $2 input and $10 output rates per million tokens run through November 30, 2026. Google’s scheduled rates afterward are $4 and $20, respectively.
4
Does Carbon Support Explicit Context Caching?
No. Carbon supports implicit caching, with a documented 4,096-token minimum, but does not support explicit context caching.
