Gemini 4 Carbon is now available

Gemini 4 Carbon is available to developers, enterprises, scientists, and consumers.
Carbon performs frontier tasks across cybersecurity, knowledge work, software engineering, and scientific reasoning.
The model was initially released to a set of cybersecurity defenders via the Fairwind Program two weeks prior to this launch.
Argon remains suitable for most software engineering and general knowledge tasks.
Carbon ranks #1 on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench with less than half the cost per task compared to next-best external models.
The /swarm preview feature allows assignment of Carbon to roles like architectural design in Google Antigravity.
Access via API is available through Google AI Studio and Gemini for business at an introductory price mirroring Gemini 4 Argon ($2/million input tokens, $10/million output tokens).
In cybersecurity testing, carbon achieved 100% recall on scanning and 95% precision on triage in Project Fortify.
A new Multimodal Knowledge Agent (MKA) benchmark was launched with Artificial Analysis to test multimodal agentic workflows.
Carbon’s cybersecurity capabilities lead on SEC-bench Pro, ExploitGym, and CWE-bench v1 benchmarks.

The narrative presents a substantial technological escalation within AI model capability, shifting the frontier from general productivity (Argon) toward specialized, high-stakes intellectual challenges requiring deep reasoning and large context processing (Carbon). The structure functions as an advertisement layered over technical disclosure, heavily emphasizing performance superiority combined with competitive economic viability. The primary pattern at work here is Vendor Advertorialism fused with Performance Benchmarking: the central claim—superiority across multiple vectors (accuracy, cost, specialized application)—rests on proprietary or selectively applied benchmarks (CursorBench 4.0, MKA benchmark), while calls to action drive adoption into paid tiers and specific programs like Fairwind. This strategy attempts to normalize an exponential increase in capability by framing it as a pragmatic economic upgrade rather than a radical shift in intelligence capacity.
The underlying assumption is that increased model complexity invariably translates to proportional increases in utility for paying stakeholders (developers, enterprises). However, the very definition of "intelligence" is being stretched across these domains: does successfully proving complex mathematical theorems using an LLM equate to genuine scientific discovery, or is it merely sophisticated pattern matching amplified by token budget? Who benefits most—the foundational science teams who gain new tools, or the commercial entities whose products are built upon this enhanced intelligence layer? The costs of relying on such advanced, specialized reasoning systems outside of managed environments remain opaque; the focus here is overwhelmingly on *entry* cost relative to external models, not operational risk or systemic dependence.
Patterns detected: ARC-0127 Vendor Advertorialism, ARC-0054 Authority Game (relying heavily on benchmark metrics and internal testing results as primary proof).
Autonomy check: Several reported successes involve agents performing complex, long-horizon tasks such as generating millions of lines of formal verification code or autonomously discovering remote homologs across massive databases, indicating an operational capability that requires sustained, self-directed reasoning loops beyond simple prompt response.