Google introduces Gemini 4 Neon, a mid-tier coding model designed for high-frequency workflows, offering Carbon-level reasoning at significantly lower costs.

On November 18, 2026, Google DeepMind officially announced the release of Gemini 4 Neon, a pivotal addition to the Gemini 4 family. Positioned as the mid-tier powerhouse, Neon is engineered to bridge the gap between the ultra-capable, long-horizon reasoning of the 'Carbon' model and the rapid, everyday execution of 'Hydrogen'.
For developers and AI engineers, this release represents a strategic shift in model deployment. Rather than forcing a choice between extreme intelligence and extreme speed, Google has provided a model specifically optimized for complex, high-frequency workflows. Neon aims to deliver frontier-level reasoning capabilities while maintaining a cost-to-performance ratio that makes agentic workflows commercially viable at scale.
Gemini 4 Neon introduces a sophisticated architecture designed for massive context handling and multi-level reasoning. The model supports a 1M token context window, allowing developers to ingest entire codebases or massive documentation sets without losing coherence. A standout feature is its support for four distinct thinking levels, with the 'xhigh' setting providing the most robust reasoning capabilities for difficult logic problems.
Beyond simple inference, Neon is built for orchestration. It is designed to act as a conductor within the Google Antigravity (/swarm) ecosystem, where it can orchestrate Gemini 4 Hydrogen instances as parallel subagents. This is facilitated by advanced context compaction techniques, which allow for long-running, stateful sessions that remain efficient even as the conversation history grows.
The benchmark data for Gemini 4 Neon places it at the top of the industry for coding-specific tasks. On the Artificial Analysis Coding Agent Index, Neon secured the #1 spot with a score of 58.6. It also demonstrates exceptional proficiency in technical environments, scoring 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. Its mathematical and scientific reasoning is equally impressive, hitting 84.0% on RiemannBench and 70.5% on Terminal-Bench Science.
Neon also dominates specialized domains, ranking #1 on the Legal Research Bench, Tax Agent Bench, and Zapier AutomationBench (56.3%). In terms of general intelligence, the Artificial Analysis Intelligence Index ranks it #3 overall at 61 (xhigh), while the Vals Index places it #2 on the Pareto frontier at 70.9%. The model's reliability is further evidenced by an Agent Arena Index of 1336 across 4,200+ sessions.
The evaluation methodology for these benchmarks involves rigorous, multi-stage testing across coding, reasoning, and agentic autonomy. By testing against both static code generation and dynamic terminal-based execution, Google has ensured that Neon's performance translates to real-world software engineering environments.
Security is a core pillar of the Neon release. The model achieves State-of-the-Art (SOTA) results on WebExploitBench v1.2 with a score of 51.26%, notably delivering this performance at half the cost of previous frontier models. Furthermore, it shows extreme robustness in security-centric benchmarks, scoring 90.5% on SEC-bench Pro and 70.7% on CWE-bench v2, which represents a 3.4% uplift over the larger Carbon model.
Safety metrics are equally stringent. Neon demonstrates a 0.0% attack success rate on the Gray Swan IPI, an agent honesty rating of 99.2%, and a remarkably low misaligned action rate of just 0.09%. For highly sensitive environments, a 'Neon Cyber' variant is also available through the Fairwind Program, specifically tailored for trusted defenders.
Google has positioned Neon to be highly competitive for high-volume users. The pricing model is designed to reward efficient development patterns, particularly through aggressive caching. At $3 USD per million input tokens, it is significantly more affordable than many frontier competitors.
The most striking feature is the 95% discount on cached inputs, which drops the cost to just $0.15 per million tokens. This makes long-context RAG (Retrieval-Augmented Generation) and continuous coding sessions extremely cost-effective. Overall, Neon is 32% cheaper per task than the Carbon model at high thinking levels and 36% cheaper than major competitors at their highest settings.
Gemini 4 Neon is the ideal choice for complex, high-frequency coding workflows. This includes automated pull request reviews, real-time pair programming, and autonomous agentic debugging. Its ability to handle specialized tasks like tax automation and legal research makes it a versatile tool for enterprise-grade RAG applications.
Additionally, the model excels in DevOps and system administration due to its high Terminal-Bench scores. It is perfectly suited for building autonomous agents that can navigate terminal environments, execute scripts, and manage complex cloud infrastructures with minimal human intervention.
Developers can begin integrating Gemini 4 Neon immediately. The model is available through several channels depending on your workflow requirements. For rapid prototyping and individual experimentation, Google AI Studio provides a seamless interface.
For enterprise-scale deployment, Neon is available via the Gemini for Business API. Users looking for advanced orchestration capabilities can access it through Google Antigravity (AI Pro/Ultra tiers). The model is also integrated directly into the Gemini app for end-user productivity.
API Pricing β Input: 3 USD per million tokens / Output: 15 USD per million tokens / Context: 1M token context window