Google Launches Gemini 4 Neon to Top Coding Agent Index

Original title: 谷歌推出 Gemini 4 Neon 模型,AA 编程 AI 智能体指数超 Claude Opus 5.5 排第一
88AI Score

Google released Gemini 4 Neon, targeting complex coding, scientific research and enterprise workflows, and claimed the top spot on Artificial Analysis' coding agent index at 58.6, ahead of Claude Opus 5.5. The model offers a 1M-token context window, four reasoning tiers, and API pricing of $3 per million input tokens and $15 per million output tokens, and can orchestrate parallel sub-agents inside the Antigravity development environment.

Full Report

IT之家 November 19 news, Google today (November 18) announced the launch of the AI model Gemini 4 Neon, focusing on complex programming, scientific research, and enterprise workflows, better balancing reasoning capability, task processing speed, and cost. It is now available to developers, enterprises, and some paid subscribers.

In terms of cost, for API calls, Gemini 4 Neon charges $3 per 1 million input tokens (IT之家 note: approximately 20 RMB at current exchange rates), $15 per 1 million output tokens (approximately 100.2 RMB at current exchange rates), and $0.15 per million cached input tokens (approximately 1 RMB at current exchange rates), and offers four reasoning levels: Low, medium, high, and xhigh. After subscribing to Google AI Pro or Ultra, users can choose to use this model in the Gemini app.

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

According to Google, Neon supports a 1 million token context window and can comprehensively process spreadsheets, PDFs, charts, audio, and video, organizing scattered information into reports and other outputs. In programming scenarios, it can troubleshoot errors across files, refactor code, and coordinate multiple AI sub-agents to execute complex tasks in the Google Antigravity development environment.

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

In terms of performance, data published by Google shows that at the highest xhigh level, Neon ranks third on the Artificial Analysis Intelligence Index with a score of 61, and ranks first on the organization's programming agent index with a score of 58.6.

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5

At the default high thinking level, its per-task cost in relevant tests is 32% lower than that of the Carbon model in the same series. These rankings and cost comparisons all correspond to specific evaluation conditions and do not represent performance on all actual tasks.

Google says that in early internal evaluations, Neon's overall task quality reached or exceeded that of the larger Gemini 4 Carbon, while task completion speed reached 3 times that of the latter.

Neon's focus is on continuously handling multi-step work, such as troubleshooting errors in large codebases, analyzing large amounts of data, and coordinating multiple AI agents to divide work and collaborate. In the Google Antigravity environment, it can act as a "dispatcher," enabling Hydrogen sub-agents to process different tasks in parallel and synchronize progress during long-running operations.

Google launches Gemini 4 Neon model, tops the programming AI agent index above Claude Opus 5.5