Google opens Gemini 4 Carbon and cuts API prices

Google releases Gemini 4 Carbon and cuts API prices

Where the model is available

Gemini 4 Argon opened to developers on October 13, and Carbon was introduced on October 28. At first, the model was available to selected cybersecurity specialists through the Fairwind program, where they could search for and fix software vulnerabilities. Google has now launched Carbon for general use.

Developers can use the model through the paid Gemini API in AI Studio and Android Studio. It has also appeared in Gemini for business, Antigravity, and specialized agents, including CodeMender and AlphaEvolve. Google AI Ultra subscribers are gradually getting access to Carbon in Antigravity, the Gemini app, and AI Mode in Search. Google is also adding the model to a program that gives participants six months of free access to the highest Antigravity usage tier; the company selects researchers to take part.

In the US, the Ultra 5x subscription costs $99.99 a month, while Ultra 20x costs $199.99. The second plan comes with higher limits. Google is also testing a feature in Antigravity that lets a team of AI agents divide up a project and review one another’s work.

Carbon is a closed model hosted on Google’s servers. The company has not announced that its weights can be downloaded or licensed for self-hosting. Launch materials describe paid access through the Gemini API but do not specify availability in Vertex AI, regions, or support terms. That may matter to teams working under specific cloud contracts.

According to Artificial Analysis, the model has a 1 million-token context window. It accepts text, images, speech, and video, and produces text. For example, teams can give it screenshots, recordings, and documentation as part of a single task.

Access to the cybersecurity version of Carbon remains limited to the Fairwind program. Anthropic draws a similar distinction: Fable and Mythos use the same base model, but Fable has stronger safeguards for cybersecurity and biology tasks. Fable is available through the Claude API and the AWS, Google Cloud, and Microsoft Azure platforms. Access to Mythos requires approval through Anthropic’s cybersecurity or biological sciences programs; Claude Enterprise customers can also use the model in the limited Claude Security product for scanning.

Token prices and task costs

The rates below apply to standard access to Carbon and cover uncached input tokens and output tokens. Both versions of Claude 5.2 also offer a fast mode that costs twice as much as the standard rate.

Until November 30, Carbon costs $2 per million uncached input tokens and $10 per million output tokens. After the introductory period, the rates will rise to $4 and $20, respectively. Google previously announced a 95% discount on input tokens read from cache. Based on the published prices, that works out to $0.10 per million such tokens during the promotion and $0.20 afterward. Teams that regularly send the same policy documents or repositories should check how often their requests are actually served from cache.

The rates listed at the beginning of the article cover Gemini 4 Carbon before and after the promotion, as well as Grok 4.7 Fast in Cursor for requests with context exceeding 256,000 and up to 500,000 tokens, and Gemini 4 Argon after its introductory period. However, that list does not provide separate prices for Argon and Grok.

Token prices do not show what a completed task costs. Artificial Analysis estimates that a run of Carbon at high reasoning effort costs an average of $7.76 per task on a general-intelligence benchmark. The estimate for Fable 5.2 at maximum effort is $26.54. At high effort, Fable costs $9.08 and scores 62 points — two more than Carbon at high reasoning effort.

At medium effort, Fable costs $3.80 and scores 56 points, while Carbon costs $6.59 and scores 58. OpenAI’s Astra at maximum effort averages $6.24 and scores 57, even though its published token rates are 2.5 times Carbon’s standard prices. Each company defines effort levels differently, so for procurement it is more useful to compare the cheapest settings that deliver the same quality, accounting for failures and retries.

The tests where Carbon outperforms competitors

In the current Artificial Analysis index, Fable 5.2 at maximum effort scores 65 points, Carbon at high effort scores 60, Claude Opus 5.5 at maximum effort scores 59, and GPT-6 Astra at maximum effort scores 57. When Carbon debuted in October, it scored 59 points. Artificial Analysis later updated its assessment of one component and recalculated the models’ results. Anthropic now leads the broad ranking of US frontier models, but Google and OpenAI are close enough that their positions depend on the task.

The strongest public evidence for Google’s pricing claims comes from Zapier’s AutomationBench test, which checks whether models can complete tasks in business applications. Carbon at high reasoning effort completed 53.88% of tasks at an estimated cost of $2.12 per task. Fable 5.2 at maximum effort scored 51.45% at $4.64.

Carbon at medium reasoning effort scored 52.97% at $1.65 per attempt. But after failed attempts are included, Zapier estimates the cost of each successfully completed task at $3.11; for Carbon at high reasoning effort, it is $3.94. For an agent updating a customer relationship management system or handling a support queue, the outcome matters: whether all systems ended up in the required state. Even the test’s best result falls short of completing every task.

Google also says Carbon ranked first in CursorBench 4.0 and Perplexity’s WANDR research benchmark, at less than half the cost of the next-best model. The launch materials do not include the scores or cost calculations behind that comparison.

Carbon is absent from the public CursorBench table reviewed by VentureBeat. Among the models listed, Fable 5.2 at maximum effort leads with 60.0% and $32.79 per task. In Anthropic’s WANDR table, Fable 5.2 scores 82.1% under a more open-ended scoring system, while Carbon’s result is not listed. Google’s claims cannot be verified from the available materials for either test.

Results in software development vary by type of work. In a direct Artificial Analysis comparison, Carbon at high reasoning effort scored 65.2% on Terminal-Bench 4.0. Fable’s best published result is 61.6%.

FrontierCode 2.0 evaluates changes in real repositories using criteria similar to those for accepting code changes. Here, Fable’s best published result is 49.1%, while Carbon at high reasoning effort scores 40.6%. Anthropic separately reports that Fable scored 69.9% on FrontierSWE v2; Astra and Carbon scored 65.5% and 61.2%.

In the Terminal-Bench-Science 0.1 scientific workflow test, Mythos and Fable lead with 75.3% and 74.6%. Astra scores 68.1%, and Carbon 64.8%. A team looking for an agent to diagnose issues through a terminal may therefore choose a different model from one seeking to increase the number of code changes that pass review.

A separate race in cybersecurity

In Google’s October evaluation, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for checking whether they could be exploited, and 71.0% on CWE-bench for fixing them. Google ran these external tests without restricting the model’s responses. The company is now directing customers who need its strongest cybersecurity capabilities to the Fairwind program.

Google says the cybersecurity version of Carbon helped Wiz find previously unknown vulnerabilities while scanning two large software-as-a-service providers. The work took less than two hours. Google did not name the companies in its launch announcement.

Anthropic says Mythos 5.2 rediscovered 89.6% of known vulnerabilities in its own test and scored 76.3% on CyScenarioBench. The company’s table does not include Carbon’s results for either measure. Anthropic also notes that the labs used different tasks for the CyScenarioBench results shown in the comparison.

Security teams will need to consider which model version they can access, then test how often it produces a reproducible finding and suggests a fix that passes internal regression tests.

Choosing a model for the work

US labs lead the broad Artificial Analysis aggregate ranking, but results vary on tests closer to specific professional tasks. In the GDPval-AA v2.1 ranking for professional work, Moonshot AI’s Kimi K3.1 Pro and Meta’s Muse Wonder Full are close to the leader, Claude Opus 5.5; the uncertainty intervals overlap.

Kimi also scores higher than any other open-weight model in the broad Artificial Analysis index. Moonshot distributes its weights under its own license. On November 9, China’s Meituan released LongCat-2.5 weights under the MIT license, while France’s Mistral offers Large 4 under its own open-weight license. For teams that need to run inference on their own infrastructure, these models meet a requirement that Carbon, Fable, and Astra do not meet when accessed through the companies’ APIs.

Google’s own products offer a way to distribute work across models. Antigravity is testing the /swarm feature: developers can assign Carbon architecture, coordination, or auditing, while lighter models handle parallel implementation tasks in isolated workspaces.

Anthropic describes a similar role for Fable: it can act as an adviser or coordinator, while the company recommends starting with Opus 5.5 for most everyday tasks. Google, in turn, recommends Argon for most software development and knowledge-work tasks.

When testing Carbon, teams should keep their current model as a point of comparison and give both systems the same tasks that already consume engineers’ time. For a complex migration, they can check whether the project builds, tests pass, and changes withstand review. In an operational workflow, they can check whether all required changes were made without touching unrelated records. They should compare model costs, total completion time, and reviewers’ time, then recalculate the cost of the same work using the rates that take effect after November.

Dataist’s take: Carbon’s lower price does not by itself mean that using it will cost less; the answer depends on task quality, retries, and review time. Public results show Carbon ahead on some tests and behind on others, while Google’s CursorBench and WANDR claims are not yet confirmed by available tables. The promotional price runs through November 30, so teams evaluating the model for their workflows should account for the rates that follow.