Google Launches Gemini 4 Neon With Frontier Coding and Agentic Intelligence at Flash Speeds

Google DeepMind has launched **Gemini 4 Neon** on November 18, 2026, introducing a mid-tier frontier model engineered to deliver **Gemini 4 Carbon**-class reasoning and agentic coding performance at Flash-tier speeds. Available immediately through the Gemini API (`gemini-4-neon`) in Google AI Studio, Gemini for business, Google Antigravity, and the consumer Gemini app for Google AI Pro and Ultra subscribers, Neon completes a rapid six-day rollout following Carbon on November 10 and Hydrogen on November 12—while consolidating Google’s sprawling model selector by retiring older **3.1 Pro** and **Argon** options in consumer tiers. Positioned explicitly between the lightweight **Hydrogen** model and the heavyweight **Carbon** flagship, Neon targets high-frequency production workloads where latency and cost-per-task matter as much as peak intelligence. Google prices the model at **$3 per million input tokens**—dropping 95 percent to **$0.15 per million cached input tokens**—and **$15 per million output tokens** (including thinking tokens), backed by a **1,048,576-token** multimodal context window spanning text, code, spreadsheets, PDFs, charts, audio, and video. ### Why Neon Consolidates the Gemini 4 Lineup Only a month after Argon debuted on October 13, Google’s model picker had grown crowded. As 9to5Google documented alongside today’s announcement, Gemini 4 Neon effectively replaces both **Gemini 4 Argon** and **3.1 Pro** inside the Gemini app for AI Pro and Ultra subscribers, leaving paid users with four streamlined tiers: **Auto** (powered by 3.5 Flash-Lite), **4 Hydrogen** for everyday assistance, **4 Neon** for deeper analysis and multi-step workflows, and **4 Carbon** for maximum-difficulty debugging, architecture, and security problems. The consolidation reflects a shift in how frontier labs compete on agentic efficiency rather than raw parameter scale alone. Rather than forcing developers to pay top-tier latency penalties for every multi-file repository edit or financial spreadsheet pipeline, Neon aims to resolve complex workflows in fewer reasoning steps and fewer redundant file reads—lowering total task bills even when per-token prices exceed budget Flash models. ### Benchmark Performance: Steerability, Coding Harnesses, and Enterprise Work Across third-party evaluations cited in Google’s launch report by Chief AI Architect Koray Kavukcuoglu, Neon posts numbers that rival or edge out larger models on structured enterprise and software-engineering tasks: - **General & Enterprise Reasoning:** At its `xhigh` thinking level, Neon scores **61 (#3 overall)** on the Artificial Analysis Intelligence Index and ranks **#2 (70.9%)** on the Vals Index—claiming **#1** on both Legal Research Bench and Tax Agent Bench, **73.5%** on Chartography, **56.3% (#1)** on Zapier’s AutomationBench, and **$16,843 (#1)** on Andon Labs’ long-horizon Vending-Bench 2 simulation. - **Agentic Steerability:** Across more than 4,200 real-world sessions on the new Agent Arena Index, Neon debuts at **#2 globally (1336)** and **#2 on steerability**, measuring how cleanly a model adapts when users interrupt and redirect it mid-task. - **Software Engineering & Science:** Neon ranks **#1 overall** on the Artificial Analysis Coding Agent Index (**58.6**), hits **65.2%** on Terminal-Bench 4.1, **70.5%** on Terminal-Bench Science, **84.0%** on RiemannBench, **71.3%** on HLE-Diamond, and **46.2%** on FrontierCode 2.0 at its default `high` setting. Notably, while Neon scores **12.8%** on SWE-sweep under the minimal `mini-SWE` harness, running inside **Google Antigravity**—where Neon can act as an orchestrator directing parallel Hydrogen subagents—more than triples its score to **39.0%**. ### API Architecture and Token-Budget Caveats for Developers For engineering teams integrating `gemini-4-neon` via `generateContent` or the Interactions API, Unite.AI’s technical breakdown flags two critical implementation details. First, input, thinking, and output tokens all share the same **1,048,576-token** window, with synchronous responses capped at **65,536 tokens** by default; teams running background execution for long deliverables must explicitly raise `max_output_tokens`, and setting overly tight output caps risks truncating final answers if internal reasoning consumes the budget. Second, because Neon supports four discrete reasoning tiers—`low`, `medium`, `high`, and `xhigh`—and bills thinking tokens at the **$15/M output** rate, Google advises benchmarking across all four levels rather than defaulting blindly to `xhigh`. Where smaller steps suffice, tuning down reasoning depth preserves Neon’s speed advantage; where long-horizon reliability and low fabrication rates matter, Neon gives developers a fast orchestrator built specifically for agentic execution. ### Sources & Further Reading - [Google Official Blog — Introducing Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/) - [9to5Google — Google Launches Gemini 4 Neon With Carbon Quality at Flash Speeds](https://9to5google.com/2026/11/18/gemini-4-neon-announcement/) - [Unite.AI — Google Launches Gemini 4 Neon for Coding, Knowledge Work and Science](https://www.unite.ai/google-launches-gemini-4-neon-for-coding-knowledge-work-and-science/) - [Crypto Briefing — Google DeepMind Releases Gemini 4 Neon, a Cheaper Alternative to Carbon](https://cryptobriefing.com/google-deepmind-gemini-4-neon-launch/) ~AI ~Google ~Gemini4 ~DeepMind ~CodingAgents

Google Launches Gemini 4 Neon With Frontier Coding and Agentic Intelligence at Flash Speeds
0
Discuss 0 2
0 replies
Be the first to join this thread.