Overview
Gemini 4 Carbon is Google DeepMind's frontier model, released on October 28, 2026, two weeks after Argon. The company positions it as its strongest model yet, with two explicit emphases: cybersecurity and long-horizon agents.
On Artificial Analysis Intelligence Index v5.0 it scores 59 and takes first place, where Argon sits at 50. On Arena it ranks first in Text Arena with 1528, second in Agent Arena with a 16.2 percent net gain across 21,900 real agent sessions, and fourth in Code Arena WebDev at 1768, an 89 Elo improvement over the previous generation.
The security numbers are published as measured without guardrails: 88.2 percent on SEC-bench Pro for vulnerability discovery, 54.8 percent on ExploitGym for exploitation and 71.0 percent on CWE-bench for patching. Alignment metrics moved too. High-severity misalignment alerts in internal real coding environments dropped by nearly 80 percent, unsupported claims roughly halved, and prompt injection success on the Gray Swan indirect benchmark went from 0.7 percent to 0.0 percent.
Key Features
- Agents that run for days: Carbon is built for persistent agentic workflows. Google says internal teams have seen it run for days working through dead ends, and that it is strongest at locating and fixing subtle bugs across large codebases and low-level systems.
- Cybersecurity across three benchmarks: 88.2 percent vulnerability discovery, 54.8 percent exploitation, 71.0 percent patching, all evaluated without guardrails. The coverage spans detection, verification and repair rather than a single task.
- Alignment numbers in the launch note: More than double the success rate at generating safe code, roughly 80 percent fewer high-severity misalignment alerts, 26 percent less overreach, 29 percent less under-service, and half the unsupported claims.
- Better web output: WebDev Arena reaches 1768, up 89 Elo from Argon. On Terminal-Bench 4.0, successful cheating falls to 0.0 percent and attempted cheating to 0.4 percent, which suggests more rule-following under evaluation.
- Staged rollout: Access opens this week to a subset of cybersecurity defenders, extends through the Fairwind Program to trusted partners, and then reaches developer, enterprise and consumer products. Carbon is also part of the US government voluntary pre-release testing program.
- Introductory pricing matches Argon: Two dollars per million input tokens and ten per million output tokens, with cached input at a 95 percent discount. After the introductory period the rates move to four and twenty dollars.
Use Cases
- Security teams validating vulnerabilities and shipping patches, using the SEC-bench Pro and CWE-bench results as a selection signal.
- Infrastructure refactors that need to run for days rather than be chopped into short sessions.
- Full-stack web delivery where front-end quality can be judged from the rendered result.
- Codebase audits that hunt subtle defects in low-level systems and turn them into fixes.
- Automated defence pipelines that chain detection, reproduction and repair.
- Regulated industries piloting frontier capability through the Fairwind Program.
Pros
- Scores 59 on Intelligence Index v5.0, currently first.
- Publishes concrete numbers across three cybersecurity benchmarks, measured without guardrails.
- First in Arena text, second in agents, backed by real session data.
- Reports five separate alignment improvements alongside capability gains.
- Introductory pricing matches Argon, with a further 95 percent discount on cached input.
- Handles work measured in days rather than turns.
Pricing
Introductory rates are two dollars per million input tokens and ten dollars per million output tokens, the same as Argon, with cached input tokens billed at a 95 percent discount. After the introductory period pricing moves to four dollars input and twenty dollars output per million tokens. Access is gated for now, starting with a subset of cybersecurity defenders and expanding through the Fairwind Program to trusted partners before developer, enterprise and consumer products. Check Google's official pricing page for the current rate card.
Summary
The interesting part of Carbon is not the leaderboard position but the choice of direction: offensive and defensive security capability, plus agents that keep running for days. The practical consequence is limited availability; the first cohort is defenders and trusted partners. Most developers will stay on Argon for now and migrate once Carbon reaches the API. Because Google published both an introductory and a standard rate, build cost models on the standard one.
Version History
- Gemini 4 Carbon released: 59 on Intelligence Index (2026-10-28): Google DeepMind released Gemini 4 Carbon two weeks after Argon. It scores 59 on Artificial Analysis Intelligence Index v5.0 for first place, against 50 for Argon. Arena ranks it first in text at 1528, second in agents with a 16.2 percent net gain, and fourth in WebDev Arena at 1768, up 89 Elo. Cybersecurity: SEC-bench Pro 88.2 percent, ExploitGym 54.8 percent, CWE-bench 71.0 percent, all without guardrails. Introductory pricing is two dollars input and ten dollars output per million tokens with 95 percent off cached input, moving to four and twenty dollars after the introductory period. Access starts with a subset of cyber defenders and Fairwind Program partners.