Google has launched Gemini 4 Neon, a model built for complex software engineering tasks that the company says is cheaper to run than its larger coding models.
Neon sits between Gemini 4 Carbon, aimed at demanding projects, and Gemini 4 Hydrogen, built for speed, giving Google a three-tier line-up for coding.
It keeps the million-token context window introduced with Gemini 3 roughly a year ago, which governs how much text and code the model can handle at once.
Neon can debug and refactor code across large repositories.
Within Antigravity, Google's development environment, it can also direct several Gemini 4 Hydrogen subagents working in parallel.
Developers using Neon through Google's application programming interface (API) will pay $3 per million input tokens and $15 per million output tokens, with cached inputs at $0.15 per million.
Tokens are the small chunks of text that AI models process and charge by.
That is more per token than Gemini 3's preview pricing of $2 per million input tokens and $12 per million output tokens for shorter prompts.
Google said Neon topped the Artificial Analysis Coding Agent Index with a score of 58.6 and scored 65.2% on Terminal-Bench 4.1.
It said Neon costs 32% less per task than Carbon at its default high-reasoning setting, and 50% to 80% less than rival models on the Vals Index and Agent Arena benchmarks.
Anand Joshi of JP Data said longer context windows often push token use beyond initial estimates once models go into production.
That could steer businesses towards smaller, cheaper models if the quality holds up, he said.
Prabhu Ram of CyberMedia Research expects a tiered approach instead, with businesses matching each model to the economics of the task rather than defaulting to premium tiers.
Joshi also warned that unclear code could keep AI confined to prototyping.
"The challenge with AI-generated code is that it could be a black box to the developer," he said.
Ram said the near-term opportunity lies in "supervised autonomy", with agents handling more complex work while developers keep oversight.
Read more about Google