SIAgentNews.com
← All news
DevelopingGemini

Google launches Gemini 4 Neon for lower-cost complex coding

The new AI model allows developers to tackle multi-file software bugs at a lower price point than larger systems while offering adjustable reasoning levels.

· First reported by eWeek

Claymation: picture for the story 'Google launches Gemini 4 Neon for lower-cost complex coding'
AI-made claymation for this story. No real people or products are shown.
39%bug repair success rate in advanced setups

The short version

  • Neon can examine code spread across multiple files to trace errors through different parts of an application.
  • The model achieved a 39% bug repair success rate when using Google's Antigravity development environment.
  • Pricing is set at $3 per million input tokens and $15 per million output tokens.
  • Developers can select from four reasoning settings to balance processing effort against task difficulty.
  • Testing shows that 'extra-high' reasoning can increase costs by 72% for only a marginal gain in intelligence.

Why it matters

For engineering teams, this provides a way to automate complex software maintenance without always paying for the most expensive models. It shifts the focus from just 'can the AI do this' to 'how efficiently can the AI do this' by allowing for granular control over compute spend.

A new tier for complex code

Google introduced Gemini 4 Neon on November 18 to address the need for high-quality programming assistance at a more accessible price. While AI can generate simple snippets of code quickly, finding and fixing bugs buried across entire software projects remains a difficult task. Neon is designed to handle these demanding assignments without requiring the high costs associated with larger models.

The model is capable of investigating why a specific feature stopped working after an update without the developer needing to point out the exact faulty file. It can look at code distributed across multiple files to follow errors through various sections of an application. This makes it a viable option for businesses looking to tackle complex software problems more efficiently.

As coding assistants begin to take on longer and more involved assignments, engineering teams are looking closely at how successfully these tools complete their work. Google aims to provide a balance between the capabilities of its top-tier systems and the need for cost-effective scaling. Developers can now choose a model that handles sophisticated logic while keeping standard prices lower.

How the model handles bugs

Neon utilizes a one-million-token context window, which allows it to process large volumes of source code and related documentation in one go. This large window is essential for understanding the interconnected nature of modern software projects. By seeing more of the project at once, the model can better understand how a change in one area might impact another.

The model's performance improves significantly when paired with the Antigravity development environment, which provides the AI with specific coding tools. In a test reported by eWeek, Neon repaired 39% of software bugs when using Antigravity, compared to just 12.8% in a simpler setup. These agents had to locate the problems themselves before they could attempt any repairs.

In one demonstration, the model used these tools to create a playable skiing simulator from scratch. This highlights the model's ability to move beyond simple text generation into functional software creation. By combining a large context window with specialized tools, the system can navigate the complexities of a full development lifecycle.

Pricing and reasoning levels

Google has set the cost for Neon at $3 for every million input tokens and $15 for every million output tokens, with additional discounts available for cached inputs. These regular rates are lower than those of Gemini 4 Carbon, which is the company's larger model. While Carbon currently has a temporary promotion making it cheaper through November 30, Neon is positioned as the standard lower-cost option for demanding tasks.

A unique feature of Neon is the inclusion of four reasoning settings that range from low to extra-high. These settings allow developers to choose exactly how much processing effort the model should devote to a specific assignment. This gives teams the ability to dial up the intelligence for very hard problems while saving money on easier ones.

However, choosing the highest setting does not always result in better performance. Analysis of independent testing showed that selecting extra-high reasoning caused average expenses to rise by about 72% and completion times to grow by 45%, while only improving the intelligence score by one point. In some specific professional document tasks, the extra-high setting actually resulted in a lower score than the high setting.

Access and implementation

The model is rolling out across several of Google's primary platforms, including the Gemini API and Google AI Studio. It is also being integrated into Android Studio and Gemini for business. This wide availability ensures that both individual developers and large enterprise teams can incorporate the tool into their existing workflows.

Eligible Google AI Pro and Ultra subscribers can access Neon through the Antigravity environment and the Gemini app. This tiered access allows different types of users to experiment with the model's capabilities depending on their subscription level. It provides a clear path for organizations to test the model's efficacy on their own specific codebases.

Because the performance of the reasoning settings can vary, Google suggests that teams benchmark the model on representative samples of their own bugs. Developers should run the same collection of bugs through different configurations to measure how many repairs pass automated testing. This data-driven approach helps teams find the sweet spot between total token spend and successful fix rates.

What to watch

  • How other developers benchmark the cost-to-success ratio of Neon compared to Claude Opus 5.5.
  • Whether more specialized coding tools are released to further improve Neon's bug repair success rate.

Try it

Run a set of known software bugs through both the 'high' and 'extra-high' reasoning settings in Gemini 4 Neon to see if the extra cost actually improves the success rate for your specific project.

Sources and links

Read the original at eWeek

First reported by eWeek. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news