Google's Gemini 4 Neon triples coding score via harness optimization
The massive performance jump suggests that the software environment managing an AI model can be just as important as the model's raw intelligence.
The short version
- Gemini 4 Neon's coding score tripled when moved from the mini-SWE harness to the Antigravity harness.
- The Antigravity harness was co-trained with Gemini and features specialized subagent management.
- Neon costs 32% less per task than the Gemini 4 Carbon model at its default high reasoning setting.
- The model is available via the Gemini API, Google AI Studio, and Gemini for Business.
Why it matters
This discovery shows that the 'wrapper' around an AI—how it reads files and manages context—can drastically change its effectiveness. For developers, it means choosing the right orchestration tool might be more impactful than swapping models.
The harness effect
Google recently revealed that its Gemini 4 Neon model achieved a significant performance leap in coding tasks without any changes to the underlying model. When tested using the mini-SWE harness, the model resolved 12.8% of the SWE-sweep coding benchmark. However, that score jumped to 39.0% when the exact same model was run inside the Antigravity harness.
This more-than-threefold improvement raises critical questions about the source of AI performance. It suggests that a model's ability to complete complex tasks may depend heavily on the harness—the software framework that manages tools, context, and execution—rather than just the model's internal weights. The results indicate that how a model interacts with a file system can be a primary bottleneck.
As thenewstack.io reported, this discrepancy highlights a growing trend in AI development where the environment is as vital as the engine. While some models like Claude Opus 5.5 saw improvements when paired with their specific provider harnesses, others like GPT-6.2 Sol actually saw their scores drop when switched to a different harness. This inconsistency proves that software optimization is not a universal fix but a specific architectural choice.
How Antigravity works
The Antigravity harness provides several specific technical advantages that help Gemini 4 Neon perform better. Unlike a generic harness, Antigravity was co-trained with Gemini to understand its specific behaviors. It is designed to handle long coding sessions by making fewer unnecessary file reads and managing context more efficiently.
One of the key features of Antigravity is its ability to run subagents in separate Git worktrees. This allows the system to isolate different parts of a coding task, ensuring that changes in one area do not interfere with another. It also includes a verification system that allows the harness to check the changes made by these subagents before they are finalized.
By managing these complex workflows, the harness takes on the heavy lifting of planning and execution. This allows the model to focus on the logic of the code rather than the logistics of navigating a large repository. The combination of co-training and specialized subagent management creates a specialized environment that maximizes the model's existing capabilities.
Cost and performance
Beyond raw performance, Google is positioning Gemini 4 Neon as a cost-effective solution for high-reasoning tasks. When set to its standard high-thinking mode, the model is 32% cheaper to run per task compared to the Gemini 4 Carbon version. This pricing efficiency is a key factor for businesses looking to scale AI-driven coding agents without ballooning their operational expenses.
The model's pricing structure is set at $3 per million input tokens and $15 per million output tokens. For users who need to process large amounts of data, cached input is available at a reduced rate of $0.15 per million tokens. These figures are based on a 1 million-token context window, providing a clear path for developers to calculate their overhead.
Google has also reported high scores for Neon on other benchmarks to validate its capabilities. It reached 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 while using its default high reasoning setting. Additionally, it secured the top spot on the Artificial Analysis Coding Agent Index with a score of 58.6, further establishing its position in the current landscape of coding tools.
Integration and scaling
Gemini 4 Neon is not intended to work in isolation but is part of a larger ecosystem of models. It is joining a model-switching system introduced with Gemini 4 Hydrogen, where different models handle different parts of a workflow. In this setup, Neon handles the planning and difficult coding problems while leaving routine work to the Hydrogen model.
This architecture allows for high-level parallelism through a feature called /swarm, which was originally introduced with Carbon. This system enables Neon to run multiple Hydrogen subagents simultaneously to complete tasks faster. While Google has not yet specified how this specific setup affects the SWE-sweep results, it points toward a multi-model strategy for complex software engineering.
Users can access Gemini 4 Neon through several different channels depending on their needs. It is available via the Gemini API, Google AI Studio, and Gemini for Business. For those who require the highest levels of performance, Google AI Pro and Ultra subscribers can access the model through the Antigravity platform.
What to watch
- How Neon performs when tested against harnesses other than Antigravity.
- The specific cost-per-task impact of the /swarm parallel subagent feature.
- Whether other frontier models see similar jumps when moved to co-trained harnesses.
Try it
Check the Gemini API pricing page to compare the cost of input and output tokens for the Neon model against your current workflow.
Sources and links
First reported by thenewstack.io. 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.