Google launches Gemini 4 Neon with a 1 million-token context window and new pricing
Article Summary
Google has introduced Gemini 4 Neon, a new model designed for complex software engineering and scientific workflows featuring a 1 million-token context window. The model is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens available at a 95% discount. Gemini 4 Neon is rolling out today via the Gemini API in Google AI Studio, for business users, and to paid Google AI Pro and Ultra subscribers.
AI Analysis
## Core Assessment
Google's release of Gemini 4 Neon sets its sights directly on long-horizon agent workflows and complex enterprise tasks, balancing top-tier benchmark rankings with aggressive token economics. Rather than just chasing peak raw reasoning scores, the release emphasizes steerability (#2 on the Agent Arena Index) and low fabrication rates alongside a 1 million-token context window. Priced at $3 per million input tokens and $15 per million output tokens, the critical architectural lever is the 95% discount for cached inputs ($0.15 per million). This pricing structure, paired with an orchestrated multi-agent pattern like the Antigravity harness, directly targets the economic pain point of agentic loops: repeatedly feeding massive repository trees, spreadsheets, or technical documentation into the context window across iterative reasoning steps.
The trade-off lives in operational complexity and harness dependence. Neon claims impressive jumps when embedded in proprietary environments—such as leaping from 12.8% on mini-SWE to 39.0% on SWE-sweep inside Google Antigravity while directing parallel sub-agents. However, achieving those claimed efficiencies requires engineering teams to actively manage four discrete "thinking levels" and build reliable context-compaction pipelines. If your architecture relies on naive API polling without prompt caching or harness-level orchestration, the model's cost and steerability advantages will quickly evaporate into the same latency and token-bloat traps seen across the rest of the frontier tier.
## Impact and Advice
- **Software Engineering Leads**: Evaluate the Gemini API in Google AI Studio for automated repository debugging and refactoring, but strictly mandate prompt caching to capture the 95% input discount on large codebases. Set default thinking levels to "high" for standard CI runs, reserving "xhigh" only for multi-file architectural edge cases.
- **Enterprise Solutions Architects**: Prototype document-heavy financial, tax, and legal automation against Neon’s long context window. Structure workflows to consolidate disparate spreadsheets and PDFs into persistent cached contexts rather than issuing repeated, uncached full-document calls across separate tool invocations.
## Bottom Line
Neon makes running massive context agents economically plausible, but the savings only materialize if your pipeline is designed to cache aggressively.
Google's release of Gemini 4 Neon sets its sights directly on long-horizon agent workflows and complex enterprise tasks, balancing top-tier benchmark rankings with aggressive token economics. Rather than just chasing peak raw reasoning scores, the release emphasizes steerability (#2 on the Agent Arena Index) and low fabrication rates alongside a 1 million-token context window. Priced at $3 per million input tokens and $15 per million output tokens, the critical architectural lever is the 95% discount for cached inputs ($0.15 per million). This pricing structure, paired with an orchestrated multi-agent pattern like the Antigravity harness, directly targets the economic pain point of agentic loops: repeatedly feeding massive repository trees, spreadsheets, or technical documentation into the context window across iterative reasoning steps.
The trade-off lives in operational complexity and harness dependence. Neon claims impressive jumps when embedded in proprietary environments—such as leaping from 12.8% on mini-SWE to 39.0% on SWE-sweep inside Google Antigravity while directing parallel sub-agents. However, achieving those claimed efficiencies requires engineering teams to actively manage four discrete "thinking levels" and build reliable context-compaction pipelines. If your architecture relies on naive API polling without prompt caching or harness-level orchestration, the model's cost and steerability advantages will quickly evaporate into the same latency and token-bloat traps seen across the rest of the frontier tier.
## Impact and Advice
- **Software Engineering Leads**: Evaluate the Gemini API in Google AI Studio for automated repository debugging and refactoring, but strictly mandate prompt caching to capture the 95% input discount on large codebases. Set default thinking levels to "high" for standard CI runs, reserving "xhigh" only for multi-file architectural edge cases.
- **Enterprise Solutions Architects**: Prototype document-heavy financial, tax, and legal automation against Neon’s long context window. Structure workflows to consolidate disparate spreadsheets and PDFs into persistent cached contexts rather than issuing repeated, uncached full-document calls across separate tool invocations.
## Bottom Line
Neon makes running massive context agents economically plausible, but the savings only materialize if your pipeline is designed to cache aggressively.