egov.mn

Gemini 4 Neon targets complex coding, research with 3x faster task completion

Google has launched Gemini 4 Neon, a new model designed to handle complex software engineering,...

This story was prepared automatically by software.

Gemini 4 Neon targets complex coding, research with 3x faster task completion

Google has launched Gemini 4 Neon, a new model designed to handle complex software engineering, scientific research, and enterprise workflows, claiming it can deliver performance comparable to its larger Gemini 4 Carbon model at three times the speed and 60 percent of the cost.

The model is designed for long-running, multi-step tasks, including debugging large codebases, analyzing extensive datasets and coordinating multiple agents to complete complex projects.

According to Google, Neon matched or exceeded Carbon in overall quality during early internal evaluations while completing tasks three times faster. The company also said the model has its lowest fabrication rate yet in internal Google Antigravity testing.

Neon scored 61 on the Artificial Analysis Intelligence Index at its highest xhigh thinking level, placing it third overall, according to Google’s reported results. It also ranked second on the Vals Index and the new Agent Arena Index for steerability, which measures how effectively a model adapts to user corrections during a task.

Faster coding and research

Software engineering is a major focus of the release. Google said Neon ranked first on the Artificial Analysis Coding Agent Index, scoring 58.6, and achieved 65.2 percent on Terminal-Bench 4.1.

The model also scored 39 percent on SWE-sweep when tested within Google’s Antigravity environment, more than tripling its 12.8 percent result on the mini-SWE harness.

These capabilities are intended to help developers work across multiple files, debug complex repositories, and make targeted changes without repeatedly processing the same information.

Neon can also act as an orchestrator in Google Antigravity, directing parallel Hydrogen subagents to handle different parts of a task. The model shares progress during long sessions and is designed to maintain performance as workflows become more complicated.

Beyond coding, Google is positioning Neon for enterprise analysis and scientific research. Its 1-million-token context window allows it to process large volumes of information across spreadsheets, PDFs, charts, audio, and video.

The model ranked first on several reported evaluations, including the Legal Research Bench, Tax Agent Bench and Zapier’s AutomationBench. It also achieved 70.5 percent on Terminal-Bench Science and 84 percent on RiemannBench, which tests mathematical reasoning.

Lower costs and stronger security

Google is also highlighting Neon’s pricing and cybersecurity capabilities. The model costs $3 per million input tokens and $15 per million output tokens. Cached input tokens cost $0.15 per million, representing a 95 percent discount.

At its default high thinking level, Google said Neon delivered a 32 percent lower cost per task than Carbon. The company also reported cost savings of roughly 50 percent to 80 percent against leading competing models on selected evaluations.

For cybersecurity, a specialized Neon version is designed to help defenders identify vulnerabilities, validate risks, and patch weaknesses in software. It scored 90.5 percent on SEC-bench Pro and 76 percent on the hard tier of AutoBaxBench, according to Google.

Access to this cybersecurity version will initially be limited to trusted defenders through Google’s Fairwind Program.

Gemini 4 Neon is rolling out through the Gemini API in Google AI Studio and Gemini for business. Google AI Pro and Ultra subscribers can also access it through the Gemini app and Antigravity.

The launch adds another model to Google’s Gemini 4 family, with Neon positioned between the faster Hydrogen model and the more powerful Carbon model for demanding, high-frequency workloads.

Source: https://interestingengineering.com/ai-robotics/google-gemini-4-neon-faster-coding-research

Share
Next storyResearch acceleration: The jagged frontier
Newsletter

Get the latest news delivered to your inbox.

Related articles