Google DeepMind has released Gemini 4 Neon, a mid-tier reasoning model in the Gemini 4 family designed for knowledge work, scientific research, and coding agents. Positioned between the higher-capacity Carbon model and the execution-focused Hydrogen model, Neon aims to offer near-frontier benchmark capabilities while reducing per-task operational costs. The model features a 1-million-token multimodal context window supporting text, audio, images, and video.
Neon introduces selectable inference thinking levels, including a default ‘high’ setting and an ‘xhigh’ mode built for complex analytical workloads. Architecturally integrated into Google Antigravity, Neon can serve as an agent orchestrator to direct parallel Hydrogen subagents on minor subtasks. Google reports that Neon cuts per-task execution costs by 32% compared to Carbon at high thinking levels, listing API prices at $3 per 1 million input tokens and $15 per 1 million output tokens.
In self-reported benchmark evaluations, Neon achieved top-tier placements across developer indexes, securing the #2 spot on the Agent Arena index ( score of 1336) and a 70.9% score on the Vals Index. On safety metrics, Google reported a 0.09% misaligned action rate on its agentic suite and zero successful prompt-injection exploits on the Gray Swan benchmark under maximum reasoning settings.
Why it matters
Provides developers a cost-optimized orchestrator model designed specifically to manage multi-agent workflows and subagent orchestration.
Accelerates the shift toward dynamic test-time compute, allowing users to scale reasoning effort via configurable thinking levels.
Pushes API pricing downward for agentic coding workloads with 32%+ task-level cost reductions compared to existing flagship models.
Source: marktechpost.com



