Google DeepMind Unveils Gemini 4 Neon
Google DeepMind has introduced Gemini 4 Neon, the latest addition to its renowned Gemini model family.
Positioned strategically between Gemini 4 Carbon and Gemini 4 Hydrogen, this innovative reasoning model is adept at managing coding agents, knowledge-driven tasks, and scientific workflows.
The developers tout it as a near-frontier solution, promising optimal performance at competitive task rates of $3 per million input tokens and $15 per million output tokens.
Summary
- Context: Offers a unified context of 1 million tokens, with a standard output capacity of 64,000 tokens, extendable up to 1 million tokens through background long decoding.
- Deployment: Exclusively runs on Google infrastructure. The model’s weights are closed, restricting local implementations or quantized versions.
- Benchmarked Performance: Achieves a score of 61 on the AA Intelligence Index, outperforming Claude Opus 5.5 (59) and GPT-6.2 Sol (57).
- Top Results: Attains an impressive 95.2% on GraphWalks (context range from 256,000 to 1 million), significantly surpassing GPT-6.2 Sol’s 80.1%.
- Lowest Performance: Registers 46.2% on FrontierCode 2.0, falling short compared to Claude Opus 5.5 at 47.3%.
- Conclusion: Most competitively priced for long-context and coding tasks at $3/$15, however, comes at a higher cost per token relative to GPT-6.2 Sol.
Understanding Gemini 4 Neon
Neon is the intermediary model in the Gemini 4 series, fulfilling roles that demand a blend of depth and efficiency.
While Carbon is tasked with the most challenging long-term projects and Hydrogen facilitates swift, everyday operations, Neon occupies a middle ground, optimizing between the two extremes.
According to the model card, Neon is an iteration derived from Gemini 4 Carbon. Google has not disclosed specifics regarding its parameter count or underlying architecture.
The model seamlessly processes diverse inputs such as text, images, audio, and video, with a knowledge cutoff extending to March 2026.
Neon’s Role in Agent Workflows
Neon is designed with multiple cognitive levels, with ‘high’ serving as the default setting. The ‘xhigh’ mode is dedicated to the most arduous engineering, analytical, and scientific tasks.
For prolonged agent engagements, Neon enhances efficiency through context compaction, minimizing redundant file read operations.
In the context of Google Antigravity, Neon functions as an orchestrator, directing parallel subagents of Gemini 4 Hydrogen on smaller, manageable tasks.
Benchmark Performance of Gemini 4 Neon

Google’s comparative analysis pits Neon against Claude Opus 5.5 and GPT-6.2 Sol. The following scores have been reported directly from Google, and have not been independently verified.
Though Neon boasts several superior metrics, it does not claim victory in every category. For instance, Opus 5.5 excels in Terminal-bench 4.1, achieving 65.8% versus Neon’s 65.2%, and similarly outperforms in FrontierCode 2.0 with a score of 47.3% compared to Neon’s 46.2%. Meanwhile, GPT-6.2 Sol takes the lead in Chartography with a score of 75.1% against Neon’s 73.5%.
Notably, on the Vals Index, Neon registers a score of 70.9%, securing the #2 position. Likewise, on the Agent Arena index, it also ranks #2 with an impressive score of 1336.
Source link: Marktechpost.com.



