Google Gemini 4 Carbon is now publicly available — and it undercuts Claude Fable 5.2 and Opus 5.5 on API price

It's the bedrock of all organic life as well as the name of Google's latest and greatest frontier model: Gemini 4 Carbon.

Now, it's finally available for most of us to use: today, Google is making Carbon generally available — nearly two weeks after the tech giant announced the model with its then-industry leading benchmarks, and granted access for selected cybersecurity defenders in its gated "Fairwind" program, where a specialized cybersecurity version with fewer guardrails is also still available.

For consumers, Carbon is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. Google’s U.S. subscription prices are $99.99 a month for Ultra 5x and $199.99 a month for Ultra 20x, which offers higher usage limits. In Antigravity, Google is also previewing a feature that lets a team of AI agents divide up a project and check one another’s work.

The wider release brings Google’s most capable AI model to the paid Gemini API in supported countries and territories and to Gemini for business. Google has yet to publish a country-by-country rollout schedule for Carbon.

VentureBeat collaborator Sam Witteveen received early access to Carbon — watch his video review on the model on YouTube.

The broader release comes a day after Anthropic launched Claude Fable 5.2 and Mythos 5.2. Fable has since taken the lead on Artificial Analysis’s general intelligence index, where Carbon debuted in first place.

Google is offering lower token prices for a model that still leads on some tests of work that enterprises can hand to agents, including a Zapier benchmark that checks whether they can complete assignments across business applications.

Carbon starts at $2 per million uncached input tokens and $10 per million output tokens. That’s one-quarter of Fable 5.2’s standard rates through Nov. 30 and half once Google raises the price to $4 and $20.

The bigger question for buyers is how many tokens, retries and human corrections it takes to finish the job. On some tasks, the model with the lower token price can cost more to run.

What Google is releasing and who can use it

Google opened Gemini 4 Argon to developers on Oct. 13 and unveiled Carbon on Oct. 28. Carbon’s October debut came with a limited rollout through Google DeepMind’s Fairwind program, where selected defenders could use it for work such as finding and fixing software vulnerabilities. Nov. 10 is the wider launch for the general-use model.

In its Nov. 10 rollout announcement, Google says developers can use the paid Gemini API through AI Studio and Android Studio. It lists API and model-picker access in Gemini for business, along with Antigravity and specialized agents such as CodeMender and AlphaEvolve. For Google AI Ultra subscribers, Carbon is rolling out in Antigravity, the Gemini app and AI Mode in Search. Google is also adding Carbon to a program that gives selected scientists six months of free access to Antigravity’s highest usage tier.

Carbon is a proprietary, hosted model. Google is distributing access through its products and API, with no downloadable weights or self-hosting license announced. The rollout materials detail paid Gemini API access. They leave Vertex AI availability, regions and support commitments unspecified, distinctions that matter for teams buying through a particular cloud contract. Artificial Analysis lists a 1 million-token context window and support for text, images, speech and video as inputs, with text output. That combination gives teams a way to bring screenshots, recordings and documentation into the same job.

Carbon for cyber remains a separate access path through Fairwind. That division resembles Anthropic’s approach. Fable and Mythos share an underlying model, with stronger cyber and biology safeguards on Fable. Fable is available through the Claude API and partner platforms including AWS, Google Cloud and Microsoft Azure. Direct access to Mythos requires approval through Anthropic’s cyber or life sciences programs, although Claude Enterprise customers can also use it through the narrower Claude Security scanning product.

What Carbon costs once the promotion ends

These are standard API rates for uncached input and output. The Carbon rates are for general-use access. Both Claude 5.2 variants also offer a faster mode at twice their standard token rates.

Google previously announced a 95% cached-input discount for Carbon. Applying that discount to the published input rates implies $0.10 per million cached input tokens during the promotion and $0.20 afterward. Teams that repeatedly send the same repository or policy documents will want to measure how often those requests actually get a cache hit.

The Artificial Analysis comparison illustrates how far the cost of a job can drift from the token price. At published standard rates, the evaluator estimates an average of $7.76 per task for Carbon at high reasoning on its index suite, compared with $26.54 for Fable at maximum effort. Fable at high effort costs $9.08 and scores 62 on the index, two points above Carbon high.

At medium effort, Fable costs $3.80 and scores 56, while Carbon costs $6.59 and scores 58. OpenAI’s Astra at maximum effort averages $6.24 and scores 57, even though its posted token rates are two and a half times Carbon’s standard rates. The effort settings are specific to each vendor. For procurement, the useful comparison is the least expensive configuration that meets the same quality target, with failures and retries included.

Where Carbon leads, and where Fable pulls ahead

The current Artificial Analysis Intelligence Index puts Fable 5.2 at 65 with maximum effort, Carbon at 60 on high, Claude Opus 5.5 at 59 on maximum and GPT-6 Astra at 57 on maximum.

Carbon’s October debut score was 59, before Artificial Analysis updated the scoring for one component and regraded models. The current rankings put Anthropic at the head of the broad U.S. frontier on that index, with Google and OpenAI close enough that the order changes with the assignment.

The strongest public evidence for Google’s cost claim comes from Zapier’s AutomationBench. Carbon high leads with 53.88% at an estimated $2.12 per task, ahead of Fable 5.2 max at 51.45% and $4.64.

Carbon medium scores 52.97% at $1.65. Once failed attempts are counted, Zapier estimates Carbon medium costs $3.11 for each passing task, compared with $3.94 for high. For a team using agents to update a CRM or handle a support queue, the ability to leave every system in the right state matters more than a low cost per attempt. Even the top score falls well short of full coverage.

Google also claims first place at less than half the cost of the next-best model on CursorBench 4.0 and Perplexity’s WANDR research benchmark. The launch materials omit the underlying score and cost figures for those comparisons.

The public CursorBench table shows Fable 5.2 max leading the listed models at 60.0% and $32.79 per task, with no Carbon row in the version VentureBeat reviewed. Anthropic’s WANDR table reports 82.1% for Fable 5.2 on its relaxed scoring measure and leaves Carbon blank. The public results reviewed for those two tests leave Google’s claims unverified.

Software engineering also splits by task.

In Artificial Analysis’s head-to-head tests, Carbon high reaches 65.2% on Terminal-Bench 4.0, ahead of Fable’s best listed run at 61.6%.

On FrontierCode 2.0, which grades changes to real repositories against the standard for merging them, Fable’s best listed run reaches 49.1% versus 40.6% for Carbon high. Anthropic separately reports 69.9% for Fable on FrontierSWE v2, ahead of Astra at 65.5% and Carbon at 61.2%.

Its scientific workflow results also favor Fable and Mythos, at 74.6% and 75.3% on Terminal-Bench-Science 0.1, versus 68.1% for Astra and 64.8% for Carbon. A team choosing an agent for terminal-heavy diagnostics could reasonably reach a different decision from one trying to get more patches through review.

The cyber race has its own scorecard

Google’s October evaluation reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for demonstrating that they can be exploited, and 71.0% on CWE-bench for patching them.

Google ran those external tests without guardrails and now points customers seeking its strongest cyber capabilities to Fairwind.

It says Wiz used Carbon for cyber to find previously unknown vulnerabilities while scanning two major software-as-a-service providers in less than two hours. Google leaves both companies unnamed in the rollout announcement.

Anthropic reports that Mythos 5.2 rediscovered 89.6% of known vulnerabilities in its internal test and scored 76.3% on CyScenarioBench. Its table supplies no Carbon result on either measure and notes that labs used different tasks for the CyScenarioBench figures shown.

A security team comparing the two has to start with which version it can access, then measure how often each model produces a reproducible finding and a patch that passes its own regression tests.

What enterprises can do with a closer race

The U.S. labs lead the current broad composite, but the picture shifts on tests closer to particular jobs. On Artificial Analysis’s GDPval-AA v2.1 professional-work leaderboard, Moonshot AI’s Kimi K3.1 Pro and Meta’s Muse Wonder Full sit close to leader Claude Opus 5.5, with overlapping uncertainty ranges.

Kimi is also the highest-scoring open-weight model on the evaluator’s broad index. Moonshot publishes the Kimi weights under a custom license. China’s Meituan released LongCat-2.5’s weights on Nov. 9 under the MIT License, and France’s Mistral offers Large 4 under its own open-weight license.

For teams that need to run inference on their own infrastructure, those models compete on a procurement requirement Carbon, Fable and Astra cannot meet through their hosted APIs.

Google’s own product design points toward mixing models. The /swarm preview in Antigravity lets developers put Carbon on architecture, coordination or audit work while assigning parallel implementation tasks to lighter models in isolated workspaces.

Anthropic describes a similar division of labor for Fable as an advisor or orchestrator, with Opus 5.5 as the starting point for most daily work. Google likewise recommends Argon for most software engineering and knowledge work.

For a Carbon pilot, that means keeping the existing model as a baseline and giving both systems the same jobs that already eat up engineering time.

A difficult migration can be judged on whether it builds, passes tests and survives review. An operations workflow can be judged on whether it makes every required change without touching unrelated records. Track model spend, elapsed time and reviewer time, then calculate what the same run would cost at December rates.

Carbon’s launch gives teams a chance to find out where Google’s extra capability earns a place in their stack before the introductory price expires.