Google has launched Gemini 4 Neon, a new AI model designed for complex software engineering, analytical work, and scientific research. The model is now rolling out across Google AI Studio, the Gemini API, Gemini for business, and the Gemini app for paid Google AI Pro and Ultra subscribers. Unlike the recently promoted Hydrogen model—now the default for general queries in Search—Neon is positioned for high-intensity, production-grade tasks requiring deep reasoning and efficiency.
Frontier performance at lower cost
Gemini 4 Neon ranks #3 on the Artificial Analysis Intelligence Index with a score of 61 at its xhigh thinking level, just behind larger, slower models. It ranks #2 on the Vals Index and achieves top positions in legal research (Legal Research Bench), tax analysis (Tax Agent Bench), and financial workflows. According to Google, Neon delivers this performance at a lower cost: 32% less per task than Gemini 4 Carbon at its default high setting, and up to 50–80% less than competing frontier models at comparable levels. Pricing is set at $3 per million input tokens ($0.15 per million cached input tokens) and $15 per million output tokens, with a 1 million-token context window.
The model excels in multi-step reasoning, structured data synthesis from mixed inputs (PDFs, spreadsheets, charts, audio, video), and long-horizon agent workflows. It ranks #1 on Zapier’s AutomationBench (56.3%) and Andon Labs’ Vending-Bench 2 ($16,843), simulating real-world business operations. For scientific and mathematical reasoning, it leads on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).
Built for secure, reliable code and cyber defence
Neon is engineered for agentic coding workflows, reducing redundant file reads, sharing progress during execution, and handling context compaction reliably over long sessions. In internal Google Antigravity testing, it achieves a 39.0% success rate on SWE-sweep—more than triple its standalone performance on the mini-SWE harness (12.8%)—demonstrating the value of its orchestration capabilities.
In cybersecurity, Neon outperforms Carbon across multiple benchmarks. It generates more secure code on AutoBaxBench, achieving state-of-the-art results on hard-tier challenges (76%). It also sets new benchmarks in penetration testing (51.26% on WebExploitBench v1.2), vulnerability detection (90.5% on SEC-bench Pro), and patching (70.7% on Collinear’s CWE-bench v2). These cyber-specific capabilities are being rolled out to trusted defenders via Google’s Fairwind Program.
Neon resists indirect prompt injection attacks entirely on the Gray Swan IPI benchmark at its highest thinking level with real-time safeguards, matching the 0.0% attack success rate of other Gemini models.
Positioning in the Gemini 4 family
Neon fills a strategic gap between Hydrogen and Carbon. Hydrogen, now the default model in AI Mode for Search globally as of 12 November, is optimised for speed and efficiency on everyday queries. Carbon remains intended for the most complex, long-running deep-thinking tasks. Neon balances depth and speed, acting as a high-judgment orchestrator that can direct parallel Hydrogen subagents in Google Antigravity.
It is also the most aligned model in the Gemini 4 series, with a 0.0% cheating rate on Terminal-Bench 4.0, 2.45% shortcutting in coding tasks (the lowest among Gemini 4 models), and a 0.09% misaligned action rate. Neon reduces under-service by 43–56% and overreach by 15–35% compared to Carbon, indicating improved discipline in agentic behaviour.
For enterprises, this means faster execution of high-frequency analytical workflows—such as financial modelling, legal research, or business automation—at predictable costs. Developers gain a production-ready model for multi-file debugging and refactoring. Pro and Ultra users can now select Neon directly in the Gemini app for advanced personal tasks.
What it means
For UK, UAE, US, and Kazakh businesses, Neon offers a viable alternative to high-cost frontier models, particularly in regulated sectors like finance and legal services. Its strong performance on structured data and lowest-to-date internal fabrication rate increase trust in automated reporting. However, access is limited to paid tiers—Google AI Pro, Ultra, or enterprise agreements—reinforcing Google’s tiered AI strategy. Organisations deeply integrated with Google Workspace gain tighter synergy with Gmail, Docs, and enterprise tools, but also deepen dependency on Google’s ecosystem.
Businesses should assess whether Neon’s cost-performance ratio justifies migration from current models, especially for workflows involving spreadsheets, compliance, or multi-step analysis. Developers can begin testing via Google AI Studio or the Gemini API. Cybersecurity teams should contact their Google Cloud account manager to request access to the Fairwind Program.
Google’s shift toward specialised models—Hydrogen for scale, Neon for precision, Carbon for depth—signals a maturing AI stack. The real test will be long-term reliability in production environments and whether independent benchmarks confirm Google’s performance and efficiency claims.