Google Launches Gemini 4 Carbon Frontier Model Ranking Number One on Artificial Analysis v5 and Leading Cybersecurity Evaluations
I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model. Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards. As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!
Context
Arriving just two weeks after Gemini 4 Argon, Gemini 4 Carbon takes first place on the Artificial Analysis Intelligence Index v5 at 59 ahead of Claude Opus 5.5 and GPT 6 Astra while setting state of the art marks across vulnerability discovery, exploitation, patching, and million token long horizon tasks.
Google DeepMind released Gemini 4 Argon two weeks earlier at an introductory API price of 2 dollars per million input tokens and 10 dollars per million output tokens. Just yesterday, Artificial Analysis overhauled its flagship evaluation suite into Intelligence Index v5, where Claude Opus 5.5 initially led at 58 ahead of GPT 6 Astra at 57 and Gemini 4 Argon at 50.
- Google Blog: Introducing Gemini 4 Carbon
- Google DeepMind Gemini 4 Carbon evaluations methodology
- Logan Kilpatrick announcement on Gemini 4 Carbon
- Google DeepMind announcement on Gemini 4 Carbon
- Elon Musk quote post calling Gemini 4 Carbon impressive
- Artificial Analysis independent breakdown of Gemini 4 Carbon
Deep dive
Google CEO Sundar Pichai, Google DeepMind CEO Demis Hassabis, and Chief AI Architect Koray Kavukcuoglu unveiled Gemini 4 Carbon on Wednesday afternoon, just fourteen days after launching Gemini 4 Argon. Carbon claims the number one spot on the newly released Artificial Analysis Intelligence Index v5 with a score of 59, surpassing Claude Opus 5.5 at 58, GPT 6 Astra at 57, and Gemini 4 Argon at 50. Across the seventeen benchmarks in Google DeepMind's evaluation table, Carbon leads fourteen, including a massive jump on long context GraphWalks from 256k to 1M tokens at 96.2 percent compared with 84.2 percent for Argon and 71.8 percent for GPT 6 Astra, 77.4 percent on Agents' Last Exam v1.1, 80.0 percent on RiemannBench, 64.8 percent on Terminal Bench Science 0.1, and 16,484 dollars on Vending Bench 2.
Cybersecurity defense and sustained long horizon software engineering are the centerpiece upgrades in Gemini 4 Carbon. Evaluated without guardrails on external cyber suites, Carbon reaches 88.2 percent on SEC bench Pro for finding real software vulnerabilities, 54.8 percent on ExploitGym for validating exploitability compared with 42.4 percent for GPT 6 Astra, and 71.0 percent on Collinear's CWE bench for root cause patching. Internally, Google reports that Carbon delivers more than double Argon's success rate on generating secure code and performed more than ten times better than another lab's frontier model when generating high quality patches for the Chrome team. On the Artificial Analysis Coding Agent Index using Antigravity CLI, Carbon scores 59, trailing only Claude Opus 5.5 at 62, while costing 7.76 dollars per Intelligence Index task at its standard 4 dollar input and 20 dollar output per million token pricing, or half that under Google's 50 percent launch discount.
Google also highlighted major alignment and anti shortcut gains. On Terminal Bench 4.0, Carbon reduced successful reward hacking to 0.0 percent and cheating attempts to 0.4 percent while cutting high severity internal misalignment alerts by nearly 80 percent and achieving a 0.0 percent attack success rate on Gray Swan's Indirect Prompt Injection benchmark.
What to expect next
Google is rolling out Gemini 4 Carbon first to vetted cybersecurity defenders this week and expanding access through its Fairwind Program and U.S. government pre release testing before opening the model to general developers, enterprises, and consumers at an introductory rate of 2 dollars per million input tokens and 10 dollars per million output tokens. By shipping a new frontier leader two weeks after Argon, Google has compressed the frontier release cycle and intensified pressure on Anthropic and OpenAI ahead of their next model updates.