Google has introduced a new reasoning neural network, Gemini 4 Neon, positioned as a tool for complex, high-intensity workflows including software development, scientific research, and corporate analytics. According to the tech giant, the model delivers cutting-edge reasoning capabilities while operating three times faster than the flagship Carbon and at a significantly lower cost. The release fills an intermediate niche in the Gemini 4 lineup, where Neon sits between the fast Hydrogen and the more powerful Carbon, which is tailored for long and complex projects.
Place in the Lineup and Operating Architecture
The primary focus of Gemini 4 Neon is on tasks where simply generating an answer is not enough: the model must analyze context, invoke tools, and execute sequential steps. A 1-million-token context window enables handling massive volumes of data—spreadsheets, PDF files, charts, audio, and video—turning bulky materials into concise summaries and reports. Within the Google Antigravity environment, the model can coordinate multiple faster Hydrogen AI agents, assigning them individual project phases and tracking progress over extended sessions.
Benchmark Results
At the maximum x-high reasoning level, the model scored 61 points in the Artificial Analysis Intelligence Index, securing third place among all participants and outperforming a number of more expensive models. Neon also ranked second in the Vals Index and the new Agent Arena Index, which evaluates how well AI executes user-directed tasks. In practical and scientific tests, the model took first place in Legal Research Bench, Tax Agent Bench, and Zapier AutomationBench, while scoring 70.5 percent in Terminal-Bench Science and 84 percent in RiemannBench, which tests mathematical capabilities. According to independent tests in FrontierCode 2.0, Neon successfully completed 46.2 percent of coding tasks, coming close to the Claude Opus 5.5 result of 47.3 percent.
Cost and Usage Economics
The model's pricing is set at $3 per million input tokens and $15 per million output tokens, while cached input tokens cost 15 cents per million—a 95 percent discount. At the standard reasoning level, task execution costs are 32 percent lower than those of Carbon, making Neon attractive for intensive workloads where both accuracy and cycle cost matter.
Cybersecurity and Availability
Google has also developed a specialized Neon version for cybersecurity, helping professionals identify vulnerabilities, assess risks, and fix code bugs. Access to this version is initially provided to verified cyber specialists through the Fairwind program. The model itself is already rolling out across the Gemini API, Google AI Studio, and Gemini for Workspace, and is also available in the Gemini app and Antigravity environment for Google AI Pro and Ultra subscribers.