Google launches Gemini 4 Carbon for cybersecurity and long-horizon tasks

A focus on defensive security

Google has officially released Gemini 4 Carbon, a frontier model designed specifically for cybersecurity and long-horizon tasks. Unlike general-purpose models that focus on single prompts, Carbon is built for workflows involving many steps and multiple tool calls. This makes it particularly useful for security teams that need to triage vulnerabilities, write patches, and test for exploitability.

To demonstrate these capabilities, Google highlighted three specific benchmarks: SEC-bench Pro, ExploitGym, and CWE-bench. These tests measure different parts of the vulnerability lifecycle, from finding a bug to proving it matters and finally fixing it. A model that performs well across all three is theoretically capable of handling a complete security workflow.

As reported by shattered.io, the model achieved 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. While these numbers show strength in identifying and patching flaws, they also highlight the complexity of the tasks. For example, the lower score on ExploitGym reflects the difficulty of proving a flaw is actually exploitable.

Benchmarking and performance

Gemini 4 Carbon currently sits at the top of the updated v5.0 Artificial Analysis Intelligence Index with a score of 59. This composite score bundles several evaluations into one number to provide a shorthand for comparing different frontier models. Because it is a composite index, a small lead often reflects minor differences across many tests rather than a dominant win in one specific area.

Some reports suggest that Carbon narrowly beats Anthropic's Claude Opus 5.5. However, Google's official documentation does not include a head-to-head table against Anthropic. Because index scores can change based on prompt wording or reasoning effort, the margin of victory may be thin. Users should look at specific configurations when comparing these two models.

The model also addresses the issue of hallucinations, which occur when an AI fabricates information. While Google mentions reduced fabrication rates in internal testing, there is no single overall hallucination percentage for Carbon. Because hallucination is hard to measure with one number, the best way to verify reliability is to run private tests on specific domain data.

Pricing and caching benefits

Google has positioned Carbon with a competitive pricing structure aimed at high-volume users. The model costs $2 per million input tokens and $10 per million output tokens. This is notably lower than the $4 and $20 pricing listed for Anthropic's Claude Opus 5.5, making Carbon a cheaper option on a per-token basis.

A significant feature for developers is the 95% discount on cached input tokens. This means that if a user resends the same system prompt, tool definitions, or repository context, the cost drops to $0.10 per million tokens. This pricing model is specifically designed to reward long-horizon tasks where agents frequently reference the same large amounts of data.

For a security pipeline sending 10 million input tokens and receiving 2 million output tokens daily, the 95% cache discount can drastically reduce costs. By keeping the context cached, a daily run could cost as little as $24.80 compared to much higher costs on models without such discounts. This makes it a strategic choice for complex, repetitive agentic loops.

Access and rollout strategy

Access to Gemini 4 Carbon is currently restricted rather than being available to the general public. Google is starting with a subset of cyber defenders before opening access to more trusted partners through its Fairwind Program. This staged rollout allows Google to monitor the model's performance and manage the risks associated with its cybersecurity capabilities.

The restricted access means that independent testers and the broader developer community cannot yet fully reproduce the benchmark numbers. This creates a gap between the official claims and independent verification. For organizations interested in the model, Google expects applicants to demonstrate specific defensive use cases and provide details on their current vulnerability backlogs.

This gated approach also helps Google build a pool of friendly testers who can provide case studies. While it slows down the pace of public scrutiny, it allows the company to tune capacity before demand spikes. For security teams, this means the path to adoption involves a vetting process rather than a simple API key signup.

The competitive landscape

Gemini 4 Carbon enters a crowded market where several labs are trading the top spot by small margins. While Google leans on price and cybersecurity specialization, Anthropic maintains advantages in context window size and cloud distribution. OpenAI is also competing with models that sit close to Carbon on the Artificial Analysis index.

The rapid pace of these releases means that leaderboard leads are often temporary. A model that holds the top spot for one month can easily be leapfrogged by a rival in the next. Because of this volatility, many buyers are looking for ways to switch between providers quickly without incurring a high migration tax for hard-coded tools.

Ultimately, the choice between models like Carbon, Opus 5.5, and others depends on specific needs like latency and data residency. While Carbon offers a strong pitch for security-focused tasks and cost-effective caching, it is one of many high-performing options. Developers may find that the best model is the one that fits their existing cloud infrastructure and specific workflow requirements.