Gemini 4 Carbon tops independent rankings, but only cyber defenders can use it

Google DeepMind has released Gemini 4 Carbon, its new flagship model, two weeks after Gemini 4 Argon. The independent lab Artificial Analysis, which runs models through its own test suite, ranked it first on Intelligence Index v5: 59 points vs. 58 for Anthropic's Claude Opus 5.5. But not everyone will get to try Carbon. The model will first go to vetted cybersecurity professionals through the Fairwind program. Google promises broad developer access "as soon as possible," without giving a date.

What Google claims about cyber defense

The company is betting on security and long multi-step tasks. Figures from its announcement:

• SEC-bench Pro, a test of finding vulnerabilities in code: 88.2%. • ExploitGym, a training environment where the model has to write a working exploit: 54.8%. • CWE-bench, fixing vulnerabilities from the CWE catalog of common weaknesses: 71.0%.

Google also boasts about honesty. On Terminal-Bench 4.0, successful cheating attempts fell to 0.0%, and the attempts themselves fell to 0.4%. Claims made without data support are roughly half as frequent as with Argon. Specs: a 1 million-token context window and the same amount on output. The model accepts text, images, video and speech, and replies in text. Cached input costs $0.20 per million tokens. Artificial Analysis lists the price as $4/$20, while Phil Schmid of Hugging Face quotes $2/$10 "with an asterisk": the 50% discount has no end date.

What independent measurements showed

• Artificial Analysis: 59 points, leading in 4 of the index's 9 tests (HLE-Diamond, TB-Science, AA-Omniscience, AutomationBench). Running the index costs $7.76 per task vs.

$10.04 for Opus 5.5 and $6.

25 for GPT-6 Astra. Carbon spends 123 thousand output tokens per task, 29% fewer than Opus. • Coding: on the Coding Agent Index, Carbon scores 59 (paired with Antigravity CLI) vs.

62 for Claude Code with Opus 5.5. FrontierCode v1.

1: 51% vs. 56%. • Office tasks: 1609 Elo on AA-Briefcase, 126 points higher than Argon.

Grok 4.7, Muse Wonder and Kimi K3.1 Pro (1678) still rank above it.

• Arena, where people vote blind on answers: first place in Text Arena (1528), second in Agent Arena, fourth in Code Arena: WebDev (1768). • Vals AI: second place on the overall index, first on the finance index and on legal search. It produces 64% more tokens than Argon.

• FrontierSWE: 61.2%, behind GPT-6 Astra (65.5), Opus 5.

5 (62.3) and Sonnet 5.5 (61.

9). SWE-sweep: a record 9.6%.

• Andon Labs: first place in Vending-Bench 2 (running a vending machine) and in Blueprint-Bench 2, where the model matched humans at spatial 3D reasoning.

There are caveats. Hallucinations rose from 4% to 11%. That is the second-best result in the tests, but the cost of an error is high in cybersecurity. The coding results were obtained in Antigravity CLI and have not been tested in other harnesses. scaling01 believes Google trails the next OpenAI and Anthropic releases by about two months. Rumors of a stronger unreleased Gemini model are unconfirmed.

The rest of the roundup

• GitHub Copilot Fusion: combines Anthropic and OpenAI models on the fly. It leads Terminal-Bench 4.0 (71), trails Opus 5.

5 in Claude Code by 2 points, but costs $5.59 per task instead of $11.42.

• H-Net v2 (Cartesia): learns to split text into tokens directly from bytes, delivering a 3–5x efficiency gain. • NVIDIA Nemotron 3.5 Transcribe: an open speech recognition model with 30B parameters (about 4B active), 40 languages, 2.

8% word error rate. • AssemblyAI-TTS-1: speech synthesis at $15 per million characters, 90 ms to first audio. • Liquid AI LFM3: a family of compact models, built around LFM3-1.

7B with a 256 thousand-token context. • Ling-3.1-flash-1019: an open MoE model with 560B parameters (25B active) and a 1 million-token context.

• Qwen4: according to a repository leak, it will get a 1 million-token context. • Anthropic S-1: a loss of about $80 billion in the first half of 2026, which Jamin Ball estimates is mainly revaluation of convertible debt and SAFE obligations, not operating losses. • OpenAI: API credits of $100–$500 per month for ChatGPT Pro subscribers.

What this means for teams

Carbon is stronger than its competitors at finding and fixing vulnerabilities and on knowledge tasks, but it remains second in programming. If you are choosing a model for security work, wait for access to open up and check the rate of false claims yourself. If you are choosing for coding, Opus 5.5 is the safer bet for now. For budgeting, what matters is not the price per token but the price per task, and here Carbon is cheaper than Opus.