What is Gemini 4 Carbon?
Gemini 4 Carbon is Google's newest frontier model, announced on October 28, 2026 for long-running coding agents and defensive security work such as finding, validating and patching vulnerabilities. It builds on Gemini 4 Argon. For now only selected cyber defenders in Google's Fairwind Program can use it; developers, businesses and consumers have no date yet.
What can you do with it?
Google positions Carbon for engineering problems that take many steps. Its teams report runs lasting days that root-cause and patch subtle bugs across large codebases and low-level systems. The typical input is a repository plus a misbehaving feature, and the intended result is a diagnosed cause and a working fix.
The second focus is defensive security across the whole cycle: finding vulnerabilities, checking which are exploitable and fixing the root cause. Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, measured without guardrails, and says the Chrome team reported Carbon performed more than ten times better than another lab's large model at generating high-quality patches. These are vendor results, not independent tests.
How it compares in Google's own table
On the Gemini page of Google DeepMind, Carbon scores 59 on the Artificial Analysis Intelligence Index v5.0, against 58 for Claude Opus 5.5, 57 for GPT-6 Astra and 50 for Gemini 4 Argon. It leads that table on Terminal-Bench 4.0 at 65.2%, but not everywhere: GPT-6 Astra and Opus 5.5 are listed higher on FrontierSWE 2, and Opus 5.5 on OSWorld-2.1. The selection of tests is Google's.
Prepare a first task while access is limited
Because Carbon is not generally available, the useful step today is an evaluation you can run on day one. Suggested workflow: pick a closed bug from your own repository whose root cause you already know, keep the failing test, and write the task as you would brief an engineer. Run it on Gemini 4 Argon now and record whether the diagnosis and patch were right.
When Carbon appears in the Gemini API, run the identical task and compare. Did it name the real cause, does the patch pass the test without touching unrelated files, and how many tokens did it spend? Google says Carbon makes about half as many unsupported claims as Argon, so also note whether it flags missing evidence instead of guessing.
Pricing and access
Access is staged. From the week of October 28, Google is giving Carbon to a subset of cyber defenders and will add more trusted partners gradually through its Fairwind Program, while it takes part in the U.S. government's voluntary pre-release testing. Google promises developers, businesses and consumers access as soon as possible, without a date. On October 30 the Gemini API pricing page listed no Carbon model or rate.
Google says the launch price will match Argon's introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper, and that $4 and $20 apply after the introductory period. The context window, output limit and Gemini app plans for Carbon have not been published.
What stands out
- Artificial Analysis Intelligence Index v5.0 score of 59, as cited by Google
- SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% (Google-reported)
- WebDev Arena 1768, in fourth place, per Google
- No Gemini API model ID or rate card published yet
Before you choose
- Only selected cyber defenders can use it for now, with no public date for others.
- No API model ID, context window or rate card published yet.
- Benchmark figures come from Google's own selection and runs.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further