Google unveils Gemini 4 Carbon, topping the Artificial Analysis index with 59 points
Google DeepMind announced Gemini 4 Carbon, its new frontier AI model, on Oct 28, two weeks after it launched Gemini 4 Argon. Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with 59 points.
Google is not releasing Carbon to the public yet. It goes first to a subset of cybersecurity defenders, with no date set for developers, businesses or consumers.
The benchmarks
Artificial Analysis, an independent evaluator, placed Carbon at its high reasoning setting ahead of the top models from Anthropic and OpenAI.
| Model | Developer | Index score |
|---|---|---|
| Gemini 4 Carbon | Google DeepMind | 59 |
| Claude Opus 5.5 | Anthropic | 58 |
| GPT-6 Astra | OpenAI | 57 |
| Gemini 4 Argon | Google DeepMind | 50 |
Google says Carbon is built for long-running engineering work. Its internal teams have run the model for days on multi-step tasks such as tracing and patching subtle bugs across large codebases. Carbon also rose 89 Elo rating points to fourth place, at 1,769, on the WebDev Arena leaderboard.
The cybersecurity results
Google names cybersecurity as the model's main gain. It reported these results, measured without safety guardrails:
- Carbon scored 88.2% on SEC-bench Pro, a test of finding real software vulnerabilities.
- It scored 54.8% on ExploitGym, a test of whether a vulnerability can be exploited.
- It scored 71.0% on Collinear's CWE-bench, a test of writing working fixes.
- On Google's internal benchmarks, it more than doubled Argon's success rate at generating secure code.
Where it trails
On Artificial Analysis's Coding Agent Index, Carbon scored 59, three points behind Claude Opus 5.5 at 62, The Decoder reported. On a factual-knowledge test, Carbon's answer accuracy rose to 58% from Argon's 48%, while its hallucination rate rose to 11% from 4%.
Google has not released a technical report or disclosed the model's parameter count, VentureBeat reported.
Access and pricing
Google says it is taking part in the US government's voluntary process for pre-release model testing. Starting the week of Oct 28, a subset of cyber defenders gets access, followed in phases by more trusted partners through Google's Fairwind Program. Google says wider release will come as soon as possible, after it gathers feedback and refines guardrails.
Carbon launches at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper. A footnote sets later prices of $4 and $20, and Google has not said when the introductory period ends, VentureBeat reported.
Who is exposed
Alphabet, Google's parent, develops Carbon through Google DeepMind and sells access to its Gemini models through Google's own products and developer tools. Anthropic and OpenAI, whose models rank second and third on the index, are private companies.
What comes next
Alphabet scheduled its third-quarter results for release after the US market close on Oct 28, hours after the announcement. Analysts expected earnings per share of $3.01 on revenue of $127.8 billion, with Google Cloud revenue up 91% to $28.9 billion, Yahoo Finance reported.
- Gemini 4 Carbon: our next frontier model
- Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot on updated, harder benchmarks — but access is limited
- Google Deepmind releases Gemini 4 Carbon just two weeks after Argon with gains across the board
- Google launches Gemini 4 Carbon AI model with cybersecurity focus
- Google to report Q3 results as investors focus on AI growth, weigh Meta Muse AI risk