Google has announced the release of its latest artificial intelligence model, Gemini 4 Neon. This advanced development is engineered for rapid execution of complex tasks across software engineering, analytical work, and scientific research.
Performance and Capabilities
Gemini 4 Neon demonstrates high efficiency, holding leading positions in reputable benchmarks. According to the Artificial Analysis Intelligence Index, the model ranks third with an ‘xhigh thinking’ score of 61, trailing only larger and slower models. On Val’s Index (vals.ai), Neon secures the second position, operating at the Pareto frontier. Furthermore, the model is recognized with a second-place ranking for controllability—its ability to adapt flexibly to user adjustments during task execution—in the new Agent Arena Index.
Neon also excels at handling multi-step instructions and exhibits the lowest fabrication rate in Google’s internal Antigravity tests.
Cost and Availability
The model is priced at $3 per million input tokens (with a 95% discount at 15 cents per million cached input tokens) and $15 per million output tokens. Gemini 4 Neon features a context window of 1 million tokens.
Gemini 4 Neon is now accessible via the Gemini API through Google AI Studio, as well as in Gemini for Business (via API and model selection). Google AI Pro and Ultra subscribers can utilize it within Google Antigravity, while Gemini Pro and Ultra users can access it through the Gemini application.
Applications Across Industries
Business, Analytics, and Science: Gemini 4 Neon can process information from diverse sources, including spreadsheets, PDF documents, audio, and video, all within its extensive context window. The model ranks second in the Val’s Index benchmark (70.9%) and holds the top spot in the Legal Research Bench and Tax Agent Bench. Neon also shows strong performance in Surge AI benchmarks, leading in categories related to finance, accounting, and spreadsheet modeling. It demonstrates high proficiency in reading complex professional diagrams (Chartography) and matches leading models in Agents’ Last Exam v1.1.
Automation and Long-Term Planning: Gemini 4 Neon ranks first in Zapier AutomationBench (56.3%) and second globally in the Arena Agent Arena Index (1336). In scientific domains, the model leads its class in Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).
Productive Code for Production Needs
Gemini 4 Neon is designed for sustained agent workflows, ensuring stable performance even in complex tasks involving numerous tools. The model efficiently handles repeated file reads, shares progress, and reliably manages context throughout extended sessions.
In code-related tasks, Neon showcases cutting-edge performance, ranking first in the Artificial Analysis Coding Agent Index (58.6%), achieving 65.2% on Terminal-Bench 4.1, and 46.2% on FrontierCode 2.0. It also effectively manages multi-file debugging and refactoring. Within Google Antigravity, using /swarm, Gemini 4 Neon generated a photorealistic 3D laboratory, a 3D ski simulator, an interactive 3D reconstruction of Alan Turing’s office, and a complex Google Sheets spreadsheet with intricate calculations.
Efficiency and Control
Google is enhancing thinking levels to provide teams with greater control. Gemini 4 Neon delivers high performance at a lower cost through optimized thinking tiers. At its highest ‘high thinking’ level, it achieves a strong score on the Artificial Analysis Intelligence Index and a 32% lower cost per task compared to Carbon. The ‘xhigh thinking’ tier enables deeper, multi-step reasoning for the most demanding tasks, while still maintaining a lower cost than competitors.
Cybersecurity Prowess
Gemini 4 Neon for cybersecurity generates more secure code than Carbon across all complexity categories on AutoBaxBench, achieving state-of-the-art results at the challenging level (76%). The model also excels on WebExploitBench v1.2 (51.26%) for vulnerability discovery, reducing costs by 50% compared to other large models. Gemini 4 Neon for cybersecurity demonstrates leading performance in identifying, validating, and patching vulnerabilities, all at a lower cost and faster speed.
Vulnerability Detection: Neon achieves top scores on SEC-bench Pro (90.5%), at roughly half the cost.
Risk Validation: On ExploitGym, the model demonstrates leading performance, using half the tokens of Carbon.
Vulnerability Patching: Neon reaches 70.7% on Collinear’s CWE-bench v2, improving on Carbon’s scores by 3.4%.
Gemini 4 Neon for cybersecurity will be available to members of the Fairwind Program in the coming days.
Reliable Operation
Gemini 4 Neon inherits security enhancements from Gemini 4 Carbon, maintaining a high level of protection against abuse and prompt injection attacks. Particular attention has been paid to:
- Reducing Abuse Risks: Neon provides robust defense against cyber and bioinformatics abuse without refusing legitimate tasks.
- Improving Model Consistency and Discipline: In internal coding tests, error rates were lower than Carbon, indicating Neon’s enhanced consistency. The model is trained to execute multi-step tasks without errors or embellishments.
- Protecting Users from Indirect Attacks: Neon demonstrates resilience against attacks where malicious instructions are embedded within web pages, emails, or documents.
Gemini 4 Neon in the Gemini 4 Family
Gemini 4 Neon is designed for complex coding, analytical reasoning, and enterprise workflows. While Carbon is suited for long-term projects and Hydrogen is built for rapid daily tasks, Neon bridges the gap by combining depth with efficiency, offering advanced reasoning capabilities for complex, high-frequency workflows faster and more affordably.
Getting Started with Gemini 4 Neon
Gemini 4 Neon begins rolling out today for developers via the Gemini API, for businesses through Gemini for Business, and for consumers within the Gemini app for Google AI Pro or Ultra subscribers. Cybersecurity professionals participating in the Fairwind Program will gain access within the next few days.









