Google opened Gemini 4 Carbon to public-facing product channels on Nov. 10, moving its highest-end model beyond a two-week trial for selected cybersecurity defenders. The release is a meaningful access and pricing change, not the model’s first appearance: paid API and enterprise users can use it through Google AI Studio and Gemini for business, while consumer access is tied to Google AI Ultra and availability is still rolling out.
The announcement positions Carbon as a model for costly, difficult assignments—long-running software projects, dense research, and other workflows that benefit from repeated reasoning and verification. Google’s description is a product claim, not a guarantee of success. The company’s model card lists benchmark scores and known limitations, while The Verge’s Nov. 10 reporting independently confirms the public rollout and introductory token prices. Neither source establishes that the product will deliver a particular result for every customer or task.
The practical question is therefore less whether Carbon is “the best” model and more who can reach it, what the announced price covers, and how to interpret performance claims that are largely based on tests selected or reported by providers. The launch also arrives with a clear access boundary: Google says consumer availability is for AI Ultra subscribers, and The Verge reports that the company had not clarified whether lower-tier subscribers or free users would receive it.
What changed on Nov. 10
Google’s announcement says Carbon is now available to developers, enterprises, scientists, and consumers. More precisely, the access paths differ by audience. Paying developers can call the model through the Gemini API in AI Studio; enterprises can use Gemini for business. Google says it is rolling Carbon out to all AI Ultra subscribers in the Gemini app, AI Mode in Search, and Antigravity, its agent-focused development environment. The model card also lists these channels, but says access varies by channel and subscription tier. Researchers may apply for a limited six-month program, subject to eligibility, capacity, and selection; this is not a general free research entitlement.
This is distinct from the earlier, narrower step. Google’s Nov. 10 post says Carbon had been made available about two weeks earlier to a set of cybersecurity defenders through the Fairwind Program. That controlled access continues as a separate path for verified defenders; it was not equivalent to consumer or broad developer availability. The title and timestamp of Google’s announcement, its concrete channel list, and The Verge’s same-day report all identify Nov. 10 as the public rollout milestone.
The distinction matters because “available” can imply universal access when it does not. API access requires a paid tier; enterprise access is through a business offering; consumer surfaces are limited to AI Ultra subscribers and are described as rolling out. The Verge explicitly says it was unclear whether Carbon would reach free users or lower-tier subscribers, so there is no basis to say that everyone can use it. Researchers’ application pathway is conditional, while cybersecurity access is for vetted defenders.
The announcement also pairs the model release with a preview of “/swarm” in Antigravity. Google describes it as a way to turn a prompt into a coordinated team of agents, assigning Carbon deeper-reasoning roles such as architectural design, coordination, and auditing, while using lighter models and verification agents in parallel. This is a preview feature and a Google description of intended operation; the Nov. 10 sources do not independently establish production reliability or outcomes at scale.
Introductory pricing—and its end date
For paid API access, Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, the same introductory rates it lists for Gemini 4 Argon. The Verge independently reports those rates and frames Carbon as competitively priced against frontier competitors. Crucially, this is not a permanent price: Google’s footnote says the public-access Carbon model’s rate changes after Nov. 30, 2026, to $4 per million input tokens and $20 per million output tokens. A reader comparing a quote or prototype must check the date and whether the relevant channel bills at API token rates.
Token prices are not the same as total task cost. A system that consumes more tokens can cost more even when its per-token rate appears modest. Google itself says Carbon uses more tokens than Argon to explore hypotheses, reason, and run extended verification loops. The Verge likewise cautions that competitors’ models may use different token counts for equivalent tasks, so per-token comparisons do not necessarily predict the invoice for a completed job. Agent orchestration, retries, context length, and the amount of generated output may further affect use; the cited launch sources do not supply a universal cost-per-task figure for customers.
Google’s announcement claims that Carbon is more than twice lower in cost per task than the next-best external model on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench. That is a company-reported comparison tied to particular benchmarks and a particular comparison set, not evidence that each live customer workflow will cost less. The model card provides AutomationBench results but not a single general-purpose accounting measure that resolves task mix, token usage, or service costs. The Verge’s reminder that benchmark results can be cherry-picked is relevant: benchmark rank and production economics are related questions, but they are not interchangeable.
A model pitched for long jobs
Google casts Carbon as a step above Argon for tasks that “test the limits” of model capability, including complex, multi-step coding and debugging and open-ended scientific research. Its framing is notable: Argon remains, in Google’s words, the ideal choice for most software engineering and knowledge work, while Carbon is meant to spend more effort on harder work. That is a product positioning statement, not a claim that Carbon should replace smaller or faster models throughout an organization.
The model card gives a technical frame for that positioning. It says Carbon is based on Gemini 4 Argon, accepts text, images, audio, and video, and has a combined input-output context window of up to one million tokens. The card describes text as the output modality and lists a March 2026 knowledge cutoff, warning that knowledge in some domains may be less current. A large context window can make it possible to submit extensive material; it does not, by itself, guarantee that every detail will be retrieved correctly, that the model’s conclusions are sound, or that a generated artifact is ready for use.
Google’s launch post gives examples meant to make “long-horizon” work concrete: agent teams assembling interactive Three.js projects from historical images or public planning documents, building a game, and modeling a musical instrument and performance hall. These are demonstrations selected and described by Google. They show the kinds of outputs the company wants users to associate with Carbon and Antigravity, but they should not be read as controlled evaluations of independent generalization, dependable engineering, or time saved for typical users.
That distinction is particularly important for agentic systems. A demonstration can show a working result while leaving unanswered how much human prompting, curation, correction, or compute was involved; whether the same workflow succeeds repeatedly; and how it behaves when inputs are incomplete or the objective has safety implications. The Nov. 10 post offers illustrative examples, not a reproducible third-party study of those questions. Teams considering use should validate their own tasks, inspect intermediate work, and preserve human review appropriate to the consequences.
How to read the performance evidence
Google reports Carbon at number one on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench. The company also says Carbon ranks second on the Multimodal Knowledge Agent benchmark, which it launched with Artificial Analysis, at more than six times lower average cost per task than the top-performing model. These claims are specific: one is a set of benchmark rankings; the other compares average cost within a newly introduced agentic multimodal test. They do not establish universal superiority across model use.
The model card, published Nov. 10, offers a broader table of results as of October 2026. It reports Carbon at 53.9 percent on Zapier AutomationBench, against 51.3 percent for Argon, 41.4 percent for GPT-6 Astra, and 42.5 percent for Claude Opus 5.5. On SWE-sweep, the card lists 9.6 percent for Carbon and 9.5 percent for GPT-6 Astra; on FrontierSWE 2, it lists Carbon at 61.2 percent, behind GPT-6 Astra at 65.5 percent and Claude Opus 5.5 at 62.3 percent. Those mixed results make a more useful reading than an unqualified “best model” label: the ordering varies by task and benchmark.
The table also lists Carbon at 65.2 percent on Terminal-Bench 4.0 (Artificial Analysis), compared with 59.1 percent for GPT-6 Astra and 59.6 percent for Claude Opus 5.5; and at 80.0 percent on RiemannBench, compared with 72.0 percent and 69.6 percent respectively. On OSWorld-2.1’s offline subset, however, the card reports 81.5 percent for Carbon and 87.2 percent for Claude Opus 5.5. Model card figures are company-published and help identify where the model performed well or less well within the listed comparisons; they are not a neutral guarantee for a buyer’s own workload.
Benchmarks depend on their tasks, scoring rules, test conditions, model versions, and tool configurations. The model card links to Google’s evaluation methodology, but the launch sources alone do not provide a complete independent replication of every listed score. The Verge notes the general risk of companies selecting favorable metrics. Readers should therefore treat a rank as evidence about performance on the specified test—not as a forecast of accuracy, reliability, speed, or cost in a different environment.
Limitations and safety context
Google’s own model card cautions that Carbon may hallucinate, that jailbreak resistance is still being improved, and that users may encounter slowness or timeouts. It adds that the model may use more tokens to maximize performance, especially at higher thinking levels. These are operationally significant caveats for an offering promoted around longer reasoning and agentic workflows: deeper or repeated processing may add latency and cost, while an apparently confident answer can still be wrong.
The card points readers to a separate safety and security report and explains that evaluation results in cybersecurity reflect performance when accessed through the Fairwind Program. Google’s framing of Carbon as having frontier cybersecurity capability should therefore be understood alongside restricted access for verified defenders, not as an invitation to assume identical access or safety controls for every public channel. The available materials do not establish that benchmark performance transfers directly to real-world incident response or that the model can safely act without human oversight.
For organizations, the sensible interpretation is a capability worth testing under governance, rather than an autonomous substitute for review. Sensitive documents, code, financial analyses, and research require checks for factual grounding, privacy and access control, security, and reproducibility. The launch materials describe channels, intended uses, and test results, but they do not provide independent evidence that Carbon improves clinical, scientific, financial, or business outcomes. Any such conclusion would go beyond the sources.
Why the availability change matters
Model launches often attract attention through a contest over benchmark leadership. This update has a more immediate practical consequence: an advanced model that was first limited to a selected cybersecurity cohort is now being made accessible across paid developer and enterprise channels, with a subscription-gated consumer rollout. The access map and announced price schedule are verifiable product facts, while claims about productivity and task economics remain dependent on workload and evaluation design.
For developers, the most consequential decision may be whether to route only the hardest requests to Carbon, as Google’s own comparison with Argon suggests, or to use it more broadly. The higher token price after Nov. 30 and the company’s note that Carbon uses more tokens support measuring end-to-end cost, not just per-token rates. Teams should compare completion quality, latency, failure recovery, tool calls, and human review time on representative jobs before broad deployment.
For consumers, “rolling out to all AI Ultra subscribers” is narrower than a general Gemini release. For researchers, the application program is conditional rather than guaranteed. For businesses, availability through a paid API or enterprise product is an access route, not an assurance that every employee or region automatically has access. These distinctions keep the announcement’s broad “now available” headline in proportion to the actual tiers and conditions described by Google and The Verge.
In short, Gemini 4 Carbon’s Nov. 10 news is its transition into a wider, still tiered release—with a published introductory API price, a fixed date for higher rates, and an accompanying agent-team preview. The case for trying it rests on measured performance for relevant tasks, not on the word “frontier.” Google’s model card supplies useful evidence and candid limitations; independent reporting verifies the rollout and price schedule while underscoring uncertainty about free access and benchmark interpretation. Those facts make Carbon a candidate for carefully scoped evaluation, not a proven shortcut to better outcomes.
Frequently asked questions
Can anyone use Gemini 4 Carbon?
No. Google says paid developers and enterprises can access it through AI Studio/API and Gemini for business, while consumer availability is rolling out to Google AI Ultra subscribers. The Verge reported that access for lower-tier subscribers and free users was unclear.
What does the API cost?
Google’s introductory rate is $2 per million input tokens and $10 per million output tokens through Nov. 30, 2026. Google says the public-access model’s rates then become $4 and $20 respectively. Token rates alone do not determine total task cost.
Is Carbon better than Argon at everything?
No such conclusion follows from the sources. Google positions Carbon for more demanding tasks and the model-card table shows mixed rankings across benchmarks. Organizations should test both models against representative workloads.
Does the launch prove improved real-world outcomes?
No. The cited materials document availability, pricing, model-card evaluations, and company demonstrations. They do not establish general productivity gains or independently verified outcomes in customers’ settings.