For two weeks, Gemini 4 Carbon was a model most people could only read about. Google had released it to a small group of cybersecurity defenders, and everyone else had to wait. On 10 November 2026 that changed: Google declared Carbon generally available, which means it is now a regular product rather than a restricted preview. If you build with the Gemini API or pay for Google AI Ultra, the question is no longer whether you can use it, but whether you should.
Who can use Gemini 4 Carbon now
Google names two groups. Paying developers and enterprises get Carbon through the Gemini API in Google AI Studio and through Gemini for business. The model ID is gemini-4-carbon, and the API documentation states that it runs on the paid tier only, so a free AI Studio project will not see it.
The second group is Google AI Ultra subscribers. For them, Carbon is rolling out in Google Antigravity, Google’s agent-based coding environment, in the Gemini app and in AI Mode in Search. “Rolling out” matters here: Google describes a gradual release, so an Ultra account may get access a little later than the announcement date.
Google also mentions two application-based routes. Scientists can apply for free six-month access to Argon and Carbon through Antigravity’s highest usage tier, subject to eligibility, capacity and selection. Security teams can still apply to the Fairwind Program for a cyber-focused version.
How access looked before 10 November
Carbon did not start as a general model. Google first released it to cyber defenders through the Fairwind Program, two weeks before the general announcement. The idea was to let security specialists use it on vulnerability work first, collect feedback and tighten guardrails, meaning the rules and filters that limit misuse, before opening it to developers, businesses and consumers. General availability is the end of that first stage, not the end of Fairwind, which continues to accept applications on a rolling basis.
What Carbon costs, and when the price rises
For now, Carbon has the same introductory price as Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. A token is a small chunk of text, roughly part of a word, and it is the unit the API bills in. According to the API documentation, this introductory rate runs through 30 November 2026. From 1 December, Carbon costs $4 per million input tokens and $20 per million output tokens.
Two details are easy to miss. First, the “thinking” tokens Carbon spends reasoning before it answers are billed as output, at the higher rate. Second, implicit caching is switched on by default for prompts of at least 4,096 tokens, and it discounts repeated input by 95 percent. If your requests keep sending the same long instructions or documents, that discount can soften the bill noticeably.
How Carbon differs from Argon
On paper, the two models look alike: both have a unified context window of 1,048,576 tokens, which is shared by the prompt, the reasoning and the answer, and both use the same built-in tools. The difference is how hard they work on each turn. Google says Carbon uses more tokens than Argon to explore thoroughly, reason deeply and run extended verification loops, in other words, rounds where the model checks and corrects its own output. The price of that thoroughness is higher token use and longer waiting times.
Google is explicit that Argon remains the recommended choice for most software engineering and knowledge work. Its documentation lists Carbon for complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and data analysis, security auditing and long, sustained agentic coding. Google’s showcase example fits that profile: from seven public city council planning documents and a single prompt, Carbon built an interactive 3D reconstruction of a Google office across 10 overnight loops.
Start with the job you need done. Here is the trade-off I would weigh before switching: the per-token price is identical to Argon only until the end of November, yet Carbon spends more tokens on every turn. The same task will therefore cost more on Carbon even this month, and from December the gap widens because the rate doubles. My inference from those numbers is simple: run a job on Argon first, and move it to Carbon only when Argon visibly fails at it, for example on a bug Argon cannot trace or a long build that needs repeated self-checks.
What Google’s honesty and safety tests claim
Google puts unusual weight on whether Carbon reports its own work honestly. On a 375-task agentic honesty benchmark, a test of whether an agent admits gaps instead of inventing results, Carbon passed 92 percent of tasks, compared with 70 percent for Argon. Google adds that Carbon declined to fill in missing experimental data 98 percent of the time.
For misuse, Google cites outside testing by Trajectory Labs, which found no new universal jailbreaks, prompts that break a model’s safeguards across many requests, during 132 hours and more than 13,000 requests. Google does not claim the problem is solved, though: it writes that robustness against prompt injection, where hidden instructions in a web page or file try to hijack an agent, remains an area for continued improvement. These are Google’s own published figures; they describe test conditions, not a promise about every real-world workflow. If you let Carbon read untrusted content on its own, a human review step still belongs in the loop.
What developers should check before switching
Moving from Argon is mostly a one-line change: swap the model ID to gemini-4-carbon. The documentation suggests keeping the default “medium” thinking level at first and testing “low”, “medium” and “high” to find the lowest level that meets your quality bar.
The operational details matter more. Synchronous responses stop at 65,536 tokens. Longer outputs require background execution with background: true and a higher max_output_tokens; the server then continues in hops of up to 64K tokens, and you join the text from each step. Hard prompts can take several minutes, so Google recommends background execution or HTTP timeouts of at least 15 minutes. Explicit context caching is not supported, and the temperature, top_p and top_k parameters are ignored if you send them. If your current code depends on any of these, test it before you point production traffic at Carbon.
