AI news ·
Google introduces Gemini 4 Neon with frontier-level intelligence for complex coding and enterprise workflows
Google DeepMind's Gemini 4 Neon ranks third globally on the Artificial Analysis Intelligence Index with a score of 61 and costs $3 per million input tokens. The model leads multiple enterprise benchmarks and delivers a 32% lower cost per task than Gemini 4 Carbon at its default setting.

Google DeepMind released Gemini 4 Neon, a new model that ranks third globally on the Artificial Analysis Intelligence Index with a score of 61 at its highest thinking level. It is priced at $3 per million input tokens and $15 per million output tokens, bringing frontier-level reasoning to complex coding, analytical, and scientific workflows at a lower cost than larger competing models.
The model is available starting today in the Gemini API through Google AI Studio, for enterprise customers via the Gemini for business API and model dropdown, and for paid Google AI Pro and Ultra subscribers in the Gemini app and Google Antigravity. Koray Kavukcuoglu, CTO of Google DeepMind, said the model is designed for "fast end-to-end task completion across complex software engineering, knowledge work, and science workflows."
Performance across enterprise and scientific benchmarks
Neon ranks second on the Vals Index with a score of 70.9% and reaches first place on both the Legal Research Bench and Tax Agent Bench. On Surge AI, it leads most benchmarks covering finance, accounting, spreadsheet modeling, and complex constraint following. It scored 73.5% on Chartography for reading dense professional charts and 75.8% on Agents' Last Exam v1.1.
The model also achieved first place on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2, a long-horizon business simulation where it generated $16,843. On scientific reasoning, Neon is best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).
Engineered for production code and agent workflows
Neon sustains performance through long-running, tool-heavy tasks and can act as an orchestrator directing parallel subagents in Google Antigravity. It ranks first on the Artificial Analysis Coding Agent Index (58.6) and scored 65.2% on Terminal-Bench 4.1. On SWE-sweep, it leads the mini-SWE harness at 12.8% and reaches 39.0% within the Antigravity environment.
In demonstrations, Neon built a photorealistic 3D lab bench with elemental visualizations, wrote a first-person WebGL2 ski simulator using real Mont Blanc elevation data and 10,200 lines of custom shaders, and recreated Alan Turing's 1936 King's College study in interactive 3D with a working Universal Turing Machine. Another test used a model swarm to turn a single image into a 100,000-cell Google Sheet running over 500,000 live formulas for edge detection, 3D depth mapping, and scene relighting.
Efficiency and cost controls
Neon offers four thinking levels that let teams adjust intelligence and token spend to specific workloads. At its default high setting, it delivers a 32% lower cost per task than Gemini 4 Carbon while maintaining a high Artificial Analysis Intelligence Index score. On the Vals Index and Agent Arena, it sits on the Pareto frontier with roughly 50% to 80% lower cost per task than top competing frontier models. The xhigh level adds deeper reasoning for the hardest workloads while still costing 36% less per task than competitors at their highest settings.
Cyber defense capabilities
Neon produces more secure code than Carbon across all difficulty tiers on AutoBaxBench, reaching 76% on the hard tier. On WebExploitBench v1.2, it is state-of-the-art at 51.26% for penetration testing at half the cost of competing large models. For vulnerability finding, it scores 90.5% on SEC-bench Pro. On ExploitGym, it matches frontier-level validation performance using half the tokens of Carbon, and it reaches 70.7% on Collinear's CWE-bench v2 for patching vulnerabilities. The cyber version rolls out to trusted defenders through the Fairwind Program in the coming days.
Safety and alignment improvements
Neon maintains Carbon-level protections against severe misuse and prompt injection while advancing model alignment. It recorded a 0.09% misaligned action rate on Google's agentic safety suite and a 99.2% pass rate on the agent honesty evaluation. On Gray Swan's indirect prompt injection benchmark, Neon resisted every attack at its highest thinking level, matching the 0.0% attack success rate of other Gemini models. It also showed reduced under-service by 43% to 56% and reduced overreach by 15% to 35% compared to Carbon across different effort levels.
Why this matters for working professionals
For developers and IT teams, Neon offers a production-ready model that handles multi-file debugging and refactoring across repositories at a lower cost per task than previous options. Customer support and operations teams can use its 1 million-token context window to synthesize information from spreadsheets, PDFs, and other documents into structured deliverables. Writers, marketers, and analysts gain a model that reads uploaded files more precisely and produces reports with a direct, natural tone. Education and healthcare professionals working with complex research material can rely on its strong performance on scientific reasoning benchmarks. The pricing structure - including a 95% discount on cached input tokens at $0.15 per million - makes frontier-level reasoning accessible for frequent, high-volume workflows across all these roles.