AI can generate code in seconds. Finding and fixing a bug buried across an entire software project is a much bigger assignment.
Google launched Gemini 4 Neon on Nov. 18, introducing a new AI model that it says can handle demanding programming work at a lower standard price than its larger systems. Developers and businesses now have another option for tackling complex software problems without paying top-tier rates.
As coding assistants take on longer assignments, engineering teams have more reason to consider how successfully these tools complete their work and how much each job costs.
Investigating bugs across entire software projects
Neon can examine code spread across several files and trace errors through different parts of an application. A developer could ask it to investigate why a feature stopped working after an update without having to identify the faulty file first.
Through Google's Antigravity development environment, the model can use coding tools to carry out more involved assignments. In one demonstration, it created a playable skiing simulator from scratch.
In a separate test, Neon repaired 39% of software bugs when working through Antigravity, compared with 12.8% using a simpler setup. Agents had to locate the problems themselves before attempting repairs, with the more advanced development environment contributing to the higher success rate.
Google's model card also places Neon close to Anthropic's larger Claude Opus 5.5 on two coding evaluations.
Coding test | What it measures | Gemini 4 Neon | Claude Opus 5.5 |
Terminal-Bench 4.1 | Completing programming and system tasks through computer commands | 65.2% | 65.8% |
FrontierCode 2.0 | Producing code changes that meet real software project requirements | 46.2% | 47.3% |
Pricing and developer access
Google priced Neon at $3 per million input tokens and $15 per million output tokens, with discounts available for cached inputs.
The AI model’s regular rates are lower than those of Gemini 4 Carbon, the company's larger model. A temporary promotion currently makes Carbon cheaper per token through Nov. 30, with standard pricing scheduled to resume Dec. 1.
A one-million-token context window allows the new system to work with substantial amounts of source code and supporting documents in a single request. Four reasoning settings, ranging from low to extra-high, let developers choose how much processing effort to devote to an assignment.
Access is rolling out through the Gemini API and Google AI Studio, as well as Android Studio and Gemini for business. Eligible Google AI Pro and Ultra subscribers can also use Neon through Antigravity and the Gemini app, depending on the service.
What eWeek found: Spending more on Neon's reasoning does not always improve results
Although Google is promoting Neon as a lower-cost option for demanding AI tasks, the amount of processing developers choose can significantly affect the final bill.
eWeek analyzed independent testing from Artificial Analysis and calculated that using extra-high reasoning increased average costs by approximately 72% and completion times by 45%, despite improving the overall intelligence score by just one point.
Artificial Analysis measurement | High reasoning | Extra-high reasoning |
Overall intelligence score | 60 | 61 |
Average cost per benchmark task | $3.74 | $6.44 |
Average time per benchmark task | 14.8 minutes | 21.5 minutes |
Professional document test score | 33.4% | 28.0% |
Figures come from Artificial Analysis's standardized evaluations. Cost and completion time reflect its benchmark workloads.
Our examination also identified an instance where additional processing delivered worse results. On the GDP.pdf evaluation of professional-document tasks, Neon scored 33.4% at high reasoning but fell to 28.0% at extra-high, a decline of 5.4 percentage points. Greater processing effort therefore produced a lower score on that assessment.
Developers and enterprise engineering teams should compare reasoning settings using assignments representative of their projects before selecting a default. One approach is to run the same collection of software bugs through both configurations, keeping the tools and acceptance checks identical.
Teams can then measure how many repairs pass automated testing and human review, alongside the total spending per approved fix.
Before defaulting to extra-high reasoning, engineering teams should benchmark both settings on a representative sample of their own bugs—using identical tools and acceptance checks to measure approved fixes against total token spend.
While maximum reasoning pays off on unusually tough problems, Neon’s advertised affordability ultimately depends on reserving that extra compute for the tasks that truly need it.
AI coding assistants are getting cheaper and more capable, giving developers new ways to tackle bugs and everyday programming tasks. Want to improve how you work with AI yourself? Our free six-minute How to Talk to AI course covers simple ways to write better prompts, with seven days of free access to other lessons. Explore courses from The Neuron Academy.


