E-Paper
Two weeks, a new AI frontier: Google Gemini 4 Carbon marks shorter upgrade cycles
Following through on the Gemini 4 Argon in just a couple of weeks, Google claims the new frontier model is better at cybersecurity and long-horizon capabilities
Prefer HTon Google
Just two weeks after Google made its Gemini 4 Argon frontier artificial intelligence (AI) model widely available for consumers and enterprises, the company already has the successor ready. It said Gemini 4 Carbon is the strongest model yet, with state-of-the-art cybersecurity and long-horizon capabilities. Availability, however, remains unclear, with the company saying Carbon will soon be available to developers, businesses, and consumers.
“Building on Gemini 4 Argon, which was launched two weeks ago, Carbon illustrates the pace at which we’re pushing the frontier forward. Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem,” said Koray Kavukcuoglu, senior vice president, Google DeepMind and chief AI architect, Google.
This week, Gemini 4 Carbon will be available to a select group of cybersecurity experts. Google said it will gradually phase in access for additional trusted partners through its Fairwind Program, under which it works with 650 partners on cybersecurity defence, fine-tuning model deployment and safety measures. These include Accenture, CrowdStrike, Snowflake and Wiz.
Compared with Gemini 4 Argon, OpenAI’s GPT-6 Astra, and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 models, Gemini 4 Argon leads on cybersecurity benchmarks. These include SEC-bench Pro (88.2% score, ahead of 85.4% for GPT-5 Astra), ExploitGYM (54.8%, ahead of 42.4% for GPT-6 Astra) and CWE-bench (71% for Gemini 4 Carbon, ahead of 68% tied between Argon and GPT-6 Astra).
The GraphWalks long-context benchmark also shows Gemini 4 Carbon scoring 96.2%, significantly ahead of Gemini 4 Astra at 84.2% and GPT-6 Astra at 71.8%.
In the Artificial Intelligence Index 50, a collective of nine AI benchmark evaluations across tasks and categories—including agentic coding, frontier agentic capabilities, and measuring factual recall and hallucination across various economically relevant domains—Gemini 4 Carbon is now at the top of the leaderboard with a score of 59, ahead of Claude Opus 5.5 (score of 58) and GPT-6 Astra (57).
Other prominent AI models released recently score significantly lower on this benchmark. Meta’s Muse Wonder Full scores 50, Kimi K3.1 Pro scores 46, and Grok 4.7 scores 34.
“Designed to take on much harder, long-horizon engineering problems, Carbon is built for sustained agentic workflows. Internally, our teams have seen Carbon run for days to persist through dead ends to tackle complex, multi-step engineering tasks like root-causing and patching subtle bugs across large codebases and low-level systems,” said Kavukcuoglu.
Google insists that Gemini 4 Carbon builds on the same multi-layered safety and security stack as Argon, which includes combining model training, real-time misuse defences, prompt-injection resilience, and chain-of-thought monitoring.
It also claims that in internal tests, Carbon reduced unsupported claims by about 2x compared with Argon and saw an uptick in collaboration as it retrieved relevant sources and surfaced discrepancies to the user before acting.
On cost, Gemini 4 Carbon will launch at the same introductory token usage price as Argon—$2 per million input tokens and $10 per million output tokens. It is unclear how long this introductory pricing will last, but after that expiry, it will cost $4 per 1M input tokens and $20 per 1M output tokens.
