Google DeepMind presented Gemini 4 Carbon, which it touts as its most powerful model to date. Designed for long‑duration agentic coding and cybersecurity defense, it tops the Artificial Analysis Intelligence Index v5.0 with a score of 59. Access, however, remains limited: the model is initially open only to a verified group of cyber defenders, not to the public API.
- Google DeepMind presents Gemini 4 Carbon, its most powerful model, built for long agentic coding and cybersecurity defense.
- Carbon leads fifteen of seventeen benchmarks, but GPT-6 Astra outperforms it on FrontierSWE.
- Access stays limited to verified cyber defenders, and the public must wait with no announced date.
- Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20.
According to Google, Gemini 4 Carbon is the second model in the Gemini 4 family, after Gemini 4 Argon. The group intends it for complex, long‑term engineering tasks. Its internal teams say they have seen Carbon work for several days on a single task, notably to pinpoint the source of subtle bugs in large codebases and low‑level systems, then fix them. Google also says it has refined the model for full‑stack web development and multimodal understanding. On WebDev Arena, it scores 89 Elo points and reaches fourth place with 1769 points.
Results published on seventeen reference tests
In its launch article, Google publishes a table of seventeen indicators, according to which Gemini 4 Carbon leads on fifteen of them, ahead of Gemini 4 Argon, GPT‑6 Astra, Claude Opus 5.5 and Claude Fable 5.1. The model was evaluated at the highest reasoning setting. Competitors’ scores are, for the most part, reported by each vendor.
In agentic coding, Carbon achieves 65.2 % on Terminal‑Bench 4.0 (AA), versus 59.6 % for Claude Opus 5.5. On FrontierSWE, however, GPT‑6 Astra outperforms the model, with 65.5 % against 61.2 %. The model reaches 77.4 % on Agents’ Last Exam v1.1. On GraphWalks, a test covering contexts of 256 000 to 1 million tokens (text units processed by the model), it records 96.2 % versus 71.8 % for GPT‑6 Astra.
The cybersecurity component is central. Google presents the model as covering the entire defense cycle: finding a vulnerability, validating it, then fixing it. Evaluations were conducted without safeguards. Carbon scores 88.2 % on SEC‑bench Pro, versus 85.4 % for GPT‑6 Astra, 54.8 % on ExploitGym versus 42.4 % for GPT‑6 Astra and 34.5 % for Opus 5.5, and 71.0 % on the CWE‑bench from Collinear, versus 68.0 % for Argon and GPT‑6 Astra. Google notes that it itself calculated the SEC‑bench Pro and ExploitGym scores, using its own evaluation environment, Antigravity. According to the company, the Chrome team judged the fixes produced by Carbon to be more than ten times better than those from a model from another lab, and Wiz reports that the model tops its web‑intrusion tests.
Google also claims gains in model behavior safety. The success rate of cheating attempts on Terminal‑Bench 4.0 drops to 0.0 %, and high‑severity misalignment alerts are said to have fallen by roughly 80 % in internal coding environments. On Gray Swan’s Indirect Prompt Injection test, the successful attack rate falls from 0.7 % for Argon to 0.0 %. Google emphasizes that these figures come from its own measurements and that no independent replication is yet available.
Gradual access, introductory pricing
Google participates in the voluntary pre‑publication testing process set up by the U.S. government. Initially, only a subset of cyber defenders has access. Other trusted partners will follow under the Fairwind program, which prioritises governments, critical‑infrastructure operators and large technology platforms. Developers, enterprises and individuals will have to wait, with no announced date. Google says it will be “as soon as possible.”
The pricing mirrors Argon’s introductory rates: $2 per million input tokens and $10 per million output tokens. After this period, prices will rise to $4 and $20. Cached inputs receive a 95 % discount.
