Gemini 4 Neon
Article URL: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/ Comments URL: https://news.ycombinator.com/item?id=50552424 Points: 316 # Comments: 304
Google has unveiled Gemini 4 Neon, an AI model that combines cutting-edge intelligence with rapid execution across complex tasks such as software engineering, knowledge work, and scientific research. Neon surpasses comparable models and ranks #3 overall with an Artificial Analysis Intelligence Index score of 61 at its high thinking level, just behind larger but slower models. On the Vals Index, Neon holds #2 place and is positioned on the Pareto frontier.
Neon emphasizes enhanced usability by securing a #2 rank in steerability on the Agent Arena Index, demonstrating its ability to adapt smoothly to mid-task user corrections. It also achieves top-tier performance in multi-step instruction following and reduces fabrication, recording the lowest rate ever in Google's Antigravity testing. Priced at $3 per million input tokens and $15 per million output tokens, Neon offers a 1-million-token context window.
The model operates seamlessly in Google's ecosystem, handling spreadsheets, PDFs, charts, audio, and videos, synthesizing diverse inputs into structured output within a single 1-million-token context window. Independent evaluations place Neon among the top models, with a 70.9% score on the Vals Index and #1 on the Legal Research Bench and Tax Agent Bench.
Neon excels in various professional domains, ranking #1 on most benchmarks related to finance, accounting, spreadsheet modeling, and complex constraint following. It achieves a 73.5% score on Chartography, reading dense professional charts, and performs on par with leading models on Agents Last Exam v1.1 (75.8%). Neon also leads in Zapier’s AutomationBench (56.3%) and Andon Labs Vending-Bench 2 ($16,843).
In scientific and mathematical reasoning, Neon demonstrates superiority in Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%). Designed for long-running agent workflows, Neon sustains performance through intricate, tool-heavy tasks, both independently and as an orchestrator directing parallel Hydrogen subagents in Google Antigravity. It reduces redundant file reads, consolidates progress during operation, and reliably manages context compaction in prolonged sessions.
Engineered for complex, tool-heavy workloads, Neon matches leading models in its class, ranking #1 overall on the Artificial Analysis Coding Agent Index (58.6) and achieving a 65.2% score on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0. For production-ready code, Neon excels at multi-file debugging and refactoring across repositories with targeted edits. In SWE-sweep, Neon leads on the mini-SWE harness (12.8%) and triples to 39.0% in Google Antigravity, showcasing the capabilities unlocked by the Antigravity harness.
Neon's four thinking levels enable teams to control intelligence and token expenditure based on workload requirements. Neon offers high performance at a lower cost per task compared to competing models on third-party benchmarks. At its default high thinking level, Neon matches a high Artificial Analysis Intelligence Index score while costing 32% less per task than Carbon.
It also occupies the Pareto frontier on both the Vals index and Agent Arena at high settings, with a 50% to 80% lower cost per task than top competing frontier models.
At the xhigh thinking level, Neon exhibits deeper, multi-step reasoning capabilities for complex engineering, analytical, and scientific tasks. Neon for cyber builds upon Gemini 4's safety and security framework, maintaining Carbon-level protections against severe misuse and prompt injection while advancing further on model alignment and agentic behavior.
Neon for cyber demonstrates state-of-the-art performance in code hardening, vulnerability finding, validation, and patching, achieving a 76% score on the hard tier of AutoBaxBench while being faster and more cost-effective than competing large models. The model is now being deployed to trusted defenders as part of the Fairwind Program.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- Gemini 4 Neon blog.google