Google has launched Gemini 4 Carbon, a new tier in its frontier model family that sits above Gemini 4 Argon. Unlike a standard version bump, Carbon is explicitly designed for tasks that require extended reasoning cycles, such as root-causing subtle bugs in large codebases or running multi-day verification loops. It is not a replacement for Argon, which remains the recommended choice for 90% of software engineering workloads, but rather a specialized tool for the hardest 10% of problems where other models typically give up.

The Test-Time-Compute Play

While Google has not disclosed Carbon's architecture or parameter count, reverse-engineering the documentation reveals a clear strategy: Carbon is a test-time-compute play. It shares the same API surface as Argon but employs a different spend profile, utilizing more tokens to explore thoroughly and reason deeply. The model supports a 1,048,576-token unified context window and introduces 'long decode' capabilities, allowing a single generation to span multiple inference calls with continuation tokens. This mechanism enables the model to resume work where it left off, though developers must manage total token budgets carefully, as costs accumulate across every hop in the chain.

Security and Governance Layers

Carbon's release highlights a trend of gated, security-focused model deployment. It launched in two stages: first to a subset of cyber defenders via the Fairwind Program on October 28, and then to general availability on November 10. The 'cyber-permissive' variant remains restricted, while the public version includes safeguards. Google reports that the restricted variant achieves 88.2% on SEC-bench Pro and 54.8% on ExploitGym, though these numbers are vendor-reported and not independently reproduced. The model also demonstrates improved alignment, with a 0.0% attack success rate on the Gray Swan Indirect Prompt Injection benchmark and a 92% pass rate on agentic honesty tests, significantly outperforming Argon's 70%.

Production Implications and Pricing

For builders, the shift to Carbon means rethinking cost models. The introductory pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026, rising to $4 and $20 thereafter. However, because Carbon is engineered to consume more tokens per task, per-token pricing is misleading; teams should budget based on cost-per-completed-task. Early adopters are using Carbon for multi-agent engineering in Google Antigravity, where it handles architectural design and auditing while paired with faster models for parallel tasks. The model also excels in knowledge work, ranking #1 on CursorBench 4.0 and #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59.

Key Takeaways

  • Carbon targets long-horizon tasks like multi-day bug hunting and security patching, not general coding.
  • It uses a test-time-compute strategy, spending more tokens for deeper reasoning and verification loops.
  • The release includes a gated 'cyber-permissive' variant for security defenders, separate from the public model.
  • Pricing is introductory through November 30, 2026, but actual costs depend on token consumption per task.
  • Independent verification of benchmarks is lacking; Google's claims of 2x Argon's success in secure code generation remain vendor-reported.

The Bottom Line

Gemini 4 Carbon is not a faster model, but a more persistent one. For teams tackling complex security audits or multi-day debugging, the ability to maintain context across long inference chains is a significant leap, provided you budget for the higher token consumption and treat the vendor-reported benchmarks with caution.