Google DeepMind released Gemini 4 Neon, a closed-weight reasoning model positioned between the flagship Carbon and lightweight Hydrogen variants, targeting coding agents, knowledge work, and scientific workflows via API-only access at $3 per 1M input and $15 per 1M output tokens. The model features a 1M-token unified context window with 64K default output, extendable to 1M via background long decode, and supports multimodal inputs with a March 2026 knowledge cutoff. Google reports Neon achieves 61 on the AA Intelligence Index, surpassing Claude Opus 5.5 (59) and GPT-6.2 Sol (57), and ranks #1 on the AA Coding Agent Index at 58.6. Key strengths include 95.2% on GraphWalks (256K–1M context) versus 80.1% for GPT-6.2 Sol and 66.8% for Opus 5.5, 39.0% on SWE-sweep (provider harness) against Opus 5.5’s 10.1%, 56.3% on Zapier AutomationBench, and 84.0% on RiemannBench. Weaknesses appear on FrontierCode 2.0 (46.2% vs Opus 5.5’s 47.3%) and Terminal-bench 4.1 (65.2% vs 65.8%). In agent workflows, Neon offers High and xHigh thinking levels, context compaction, and orchestration of parallel Hydrogen subagents via Google Antigravity. Google claims 32% lower per-task cost than Carbon at High thinking and 50–80% savings versus rivals on Vals Index (70.9%) and Agent Arena (#2, 1336). Safety metrics show 0.09% misaligned action rate, 99.2% agent honesty pass rate, 0.0% prompt injection success on Gray Swan, and a 2.45% coding shortcut rate. All benchmarks are self-reported by Google.
Read original
marktechpost