11/17/2026, 09.50
by Danielecovers artificial intelligence and infrastructure
Condividi su Facebook Condividi su Twitter Condividi su Pinterest Condividi su Telegram Condividi su WhatsApp

Google Gemini 4 Carbon: New Frontier Model Claims AI Top Spot

Google unveils Gemini 4 Carbon, targeting cybersecurity and long-horizon engineering. It leads the Artificial Analysis Index but remains in limited release.
Google Gemini 4 Carbon: New Frontier Model Claims AI Top Spot
Key points
  • Gemini 4 Carbon takes the #1 spot on Artificial Analysis Intelligence Index v5.0 with 59 points.
  • The model specializes in cybersecurity, vulnerability patching, and multi-day autonomous engineering tasks.
  • Access is currently restricted to the Fairwind Program for selected cybersecurity defenders.
  • Introductory pricing matches Gemini 4 Argon, though rates are expected to double after the initial period.

The race for frontier AI supremacy has entered a phase of hyper-acceleration. Only two weeks after the broad release of Gemini 4 Argon, Google has unveiled Gemini 4 Carbon, a model designed to push the boundaries of autonomous software engineering and cybersecurity. This rapid iteration highlights a volatile market where the title of world-leading AI shifts in a matter of days, as Google, OpenAI, and Anthropic trade blows across increasingly difficult benchmarks.

A new leader on the Intelligence Index

According to independent evaluator Artificial Analysis, Gemini 4 Carbon has claimed the top position on the Intelligence Index v5.0. With a score of 59 points, Carbon edges out Anthropic's Claude Opus 5.5, which scored 58, and OpenAI's GPT-6 Astra, which followed with 57. This shift is particularly notable because it utilizes high reasoning capabilities to outperform its predecessors and rivals in complex cognitive tasks.

Google's own internal data supports this claim, showing Carbon leading in 15 out of 17 key benchmarks. However, the dominance is not absolute. OpenAI still maintains an edge in the FrontierSWE 2 software engineering test, while Anthropic remains the leader in OSWorld-2.1, a benchmark specifically measuring a model's ability to operate a computer interface. For the global enterprise, this suggests that while Carbon is the most intelligent generalist, specialized tasks still favor specific competitors.

Targeting the cybersecurity frontier

Unlike previous releases that focused on general productivity, Carbon is positioned as a specialized tool for secure, autonomous engineering. Google is leaning heavily into the security vertical, claiming significant gains in finding vulnerabilities and generating patches. The model has posted impressive results in specialized tests: 88.2% on SEC-bench Pro for vulnerability discovery, 71.0% on CWE-bench for patching, and 54.8% on ExploitGym for exploitation testing.

The goal is to move beyond simple code suggestions toward a system that can autonomously secure an entire codebase. Internal reports indicate that Carbon has more than doubled the secure-code success rate of the previous Argon model. Furthermore, the model shows a high resistance to adversarial attacks, blocking every simulated indirect prompt injection in the Gray Swan benchmark and attempting to cheat on only 0.4% of Terminal-Bench 4.0 questions, with zero successful attempts.

The shift toward long-horizon autonomy

One of the most significant architectural leaps in Gemini 4 Carbon is its capacity for long-horizon tasks. While most LLMs operate in short bursts of interaction, Google reports that Carbon can remain active for days at a time to tackle multi-step engineering problems. This allows the AI to persist through dead ends, root-cause subtle bugs across massive codebases, and handle low-level system migrations without human intervention at every step.

This agentic capability is reflected in its performance on Terminal-Bench 4.0, where it jumped eight points over Argon. By typing commands directly into a computer terminal to install software or fix broken code, Carbon is moving closer to the ideal of an autonomous AI engineer. It has also climbed to fourth place on the WebDev Arena, a leaderboard based on human votes for AI-generated websites.

Limited access and the Fairwind Program

Despite the aggressive benchmarks, most businesses cannot yet access Gemini 4 Carbon. Google is employing a cautious rollout strategy, granting initial access only to a curated group of cybersecurity defenders through the Fairwind Program. This approach allows Google to iterate on guardrails and gather feedback from trusted partners before a general release.

This restricted launch is partly a response to the risks associated with high-capability models. As noted by industry observers, incidents like PHP2Shell demonstrate the potential havoc frontier models can wreak if deployed without sufficient safeguards. Google has confirmed it is participating in the U.S. government's voluntary process for pre-release model testing to mitigate these risks. Until the broad release, which Google promises will happen as soon as possible, Gemini 4 Argon remains the primary model available to the public.

Pricing structures and hidden tradeoffs

To maintain competitiveness, Google has launched Carbon with introductory pricing identical to Argon. The current rates are set at $2 per million input tokens and $10 per million output tokens, with a substantial 95% discount for cached inputs. However, this is a temporary window. A footnote in the announcement reveals that prices will eventually rise to $4 per million input tokens and $20 per million output tokens, though a specific deadline for this increase has not been provided.

Independent testing reveals that while Carbon improves substantially over Argon in coding and knowledge work, it trails several rivals on professional tasks and records a higher hallucination rate than Argon on a factual-knowledge test.

This highlight a critical trade-off for CTOs: the increase in reasoning and coding power comes with a slight decrease in factual reliability compared to its predecessor. For enterprises, the choice between Argon and Carbon may depend on whether they prioritize absolute factual accuracy or high-level engineering autonomy.

Global implications for US and UK enterprises

For companies in the USA and UK, the arrival of Gemini 4 Carbon signals a shift in how AI will be integrated into the DevOps pipeline. The ability of a model to run for days on a single engineering task suggests that AI is moving from a copilot to an autonomous agent. This will likely accelerate the adoption of AI-driven security auditing, reducing the reliance on manual penetration testing for initial vulnerability discovery.

From a regulatory perspective, the U.S. government's involvement in the pre-release testing of Carbon suggests that frontier models are now viewed as critical infrastructure. UK firms, which have historically leaned toward a more flexible AI regulatory environment, will find Carbon's capabilities particularly useful for scaling software exports and maintaining high security standards without exponentially increasing headcount. However, the limited access via the Fairwind Program means that for the immediate future, the competitive advantage will reside with those few firms already embedded in Google's trusted partner ecosystem.

FAQ

How does Gemini 4 Carbon differ from Gemini 4 Argon?

Carbon is more intelligent, ranking higher on the Artificial Analysis Intelligence Index, and specializes in long-horizon engineering tasks and cybersecurity, whereas Argon is the current generally available model.

Who can currently use Gemini 4 Carbon?

Access is currently limited to selected cybersecurity defenders through the Fairwind Program.

What are the pricing details for the new model?

Introductory prices are $2 per million input tokens and $10 per million output tokens, but these are expected to increase to $4 and $20 respectively.

Does Carbon outperform all other models in every category?

No. While it leads the Intelligence Index, GPT-6 Astra still leads in some software engineering tests (FrontierSWE 2) and Claude Opus 5.5 leads in computer-use evaluations (OSWorld-2.1).


Sources: Venturebeat, Blog, Nexacurrent ·

Hai una domanda su questo dossier?

Scrivila qui: Susanna, l assistente AI di glacom, ti risponde via email con un approfondimento gratuito.

Nessuna consulenza personalizzata (finanziaria, legale o medica): solo informazione e fonti. Email usata solo per rispondere.

oppure scrivile su: WhatsApp · Telegram · SimpleX · Delta Chat · Email

Condividi su Facebook Condividi su Twitter Condividi su Pinterest Condividi su Telegram Condividi su WhatsApp
Printable version
CLOSE X
See also
Anthropic's Claude AI Accelerates Drug Development Pipelines
Anthropic reveals how Claude AI is slashing drug batch review time by 89% and designing thousands of antibodies through partnerships with six biotech …
17/11/2026 13:19
GitHub Copilot Fusion: Slashing AI Coding Costs via Orchestration
GitHub debuts Copilot Fusion, an orchestration engine that mixes OpenAI and Anthropic models to match top performance at nearly half the per-task cost…
17/11/2026 11:06
Microsoft Opens Copilot Harness in Foundry Agent Service for AI Agents
Microsoft launches the Copilot harness in Foundry Agent Service, providing a managed orchestration layer to build complex, multi-model AI agents for e…
17/11/2026 07:52
Bitly Data: Short Links Aren't the Main Cause of Conversion Loss
New Bitly research reveals that while generic short links can cause hesitation, the real conversion killer is directing social traffic to generic home…
16/11/2026 19:05
AI Bot Attack Distorts UK Government Vaping Consultation
A scientific study reveals that 76% of responses in a UK government consultation on vaping were AI-generated bots attempting to block a disposable vap…
16/11/2026 16:42