inferwire
/
AI·3 min read

Google Debuts Gemini 4 Neon for Agentic Coding and Tasks

Google released Gemini 4 Neon, balancing speed and reasoning across multi-step software engineering, cyber defense, and enterprise workflows.

TL;DR

  • Google introduced Gemini 4 Neon, a model optimized for fast task completion across software engineering, analytical, and scientific workflows [^1].
  • The model balances reasoning speed and cost, reaching generation speeds of 127 tokens per second in OpenRouter testing [^2].

Background

AI developers build model families to balance execution speed, reasoning depth, and operational costs. While large flagship models handle complex long-horizon tasks, their slow response speeds and high token prices can limit real-time applications. To address these trade-offs, providers deploy mid-tier models optimized for multi-step agentic execution, structured data synthesis, and rapid tool use across enterprise environments.

What happened

Google DeepMind introduced Gemini 4 Neon, positioning the model between its long-horizon Carbon model and fast everyday Hydrogen model [^1]. Neon features a 1 million-token context window priced at $3 per million input tokens, $0.15 per million cached input tokens, and $15 per million output tokens [^1]. In evaluations, Neon scored 61 at its xhigh thinking level on the Artificial Analysis Intelligence Index, ranking #3 overall [^1]. On the Agent Arena Index, it debuted at #2 globally with a score of 1336 across more than 4,200 real-world agent sessions [^1]. Early serving via OpenRouter recorded generation speeds of 127 tokens per second for Neon, compared to 46 tokens per second for Gemini 4 Carbon [^2].

For software engineering, Neon ranked #1 overall on the Artificial Analysis Coding Agent Index with a score of 58.6 [^1]. It achieved 65.2% on Terminal-Bench 4.1, scored 12.8% on the mini-SWE harness, and reached 39.0% on SWE-sweep in Google Antigravity [^1]. In cyber defense benchmarks, Neon reached 90.5% on SEC-bench Pro for finding vulnerabilities and scored 51.26% on WebExploitBench v1.2 [^1]. Google is rolling out Neon across the Gemini API, Google AI Studio, Gemini for business, and subscriber tiers in Google Antigravity and the Gemini app [^1].

Why it matters

Neon targets a specific bottleneck in agentic workflows: balancing high-judgment reasoning with execution speed and token cost. Long-running AI agents that refactor multi-file repositories or run continuous threat analysis consume massive context histories. Neon offers four thinking levels [^1]. At its default high setting, Neon provides a 32% lower cost per task than Carbon [^1]. At its xhigh thinking level, it costs 36%+ less per task than competing models at their highest setting [^1].

The release also reflects growing emphasis on agent safety, alignment, and resistance to prompt injection. Neon recorded a 0.09% misaligned action rate on Google's agentic safety suite and achieved a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark at its highest thinking level [^1]. Statistical modeling projects Neon's performance on the Agent Duel leaderboard to reach an Elo of 1,066, landing close to Opus 5.5 at 1,078 [^2].

Practical example

Say Alex manages a security team auditing enterprise web applications for unpatched code flaws.

Instead of manually reviewing repositories, Alex configures an agentic workflow using Gemini 4 Neon through Google Antigravity [^1]. When supplied with files across an enterprise project, Neon leverages its 1 million-token context window to locate vulnerabilities, achieving 90.5% on SEC-bench Pro [^1].

The agent checks exploitability, utilizing Neon's cyber tools to validate risks on ExploitGym while using half the tokens of Carbon [^1]. When fixing identified flaws, Neon patches code across repositories with targeted edits, achieving a 70.7% score on CWE-bench v2 [^1]. Throughout the multi-step session, Neon maintains safeguards, refusing 0% of tasks on Terminal-Bench-Science while keeping its misaligned action rate down to 0.09% [^1].

Alex receives verified patches for critical vulnerabilities faster, reducing remediation time without risking unvetted code execution [^1].

Sources

  1. [1]Google — Introducing Gemini 4 Neon
  2. [2]micjm — Gemini 4 Neon looks like an Opus 5.5 competitor
  3. [3]Hacker News — Gemini 4 Neon