Google DeepMind launches Gemini 4 Neon, a fast, low‑cost frontier model for code, enterprise and science
Google DeepMind unveiled Gemini 4 Neon, a frontier‑level model priced at $3 per million input tokens.
Source: Google DeepMindGemini 4 Neon arrives as the latest model from Google DeepMind, promising frontier‑level intelligence while cutting cost per task. Launched on November 18, 2026, the model is positioned for fast, end‑to‑end completion of software engineering, knowledge‑work and scientific workflows, and it is immediately available through the Gemini API, Google AI Studio, and the Gemini app for Pro and Ultra subscribers. According to the DeepMind blog, Neon’s blend of speed, context size and safety marks a notable shift for production‑grade AI.
What is Gemini 4 Neon and what are its headline capabilities?
Neon is built around a 1 million‑token context window, allowing it to ingest and synthesize large multimodal inputs such as spreadsheets, PDFs, charts, audio and video in a single session. The model is described as “frontier‑level” intelligence that can orchestrate complex tool‑heavy tasks, act as an Orchestrator for parallel Hydrogen subagents, and maintain performance over long‑running agent workflows. Safety‑related improvements carry forward Gemini 4 Carbon’s protections against misuse and prompt injection while advancing alignment for agentic behavior.
How does Neon rank on benchmark evaluations?
Independent evaluations place Neon near the top of several leaderboards. On the Artificial Analysis Intelligence Index it scores 61, ranking #3 overall and #2 on the Vals Index (70.9%). It lands on the Pareto frontier of the Agent Arena Index (#2 with a score of 1336 across 4,200+ sessions) and is #2 for steerability on the new Agent Arena Index. In enterprise benchmarks it hits #1 on the Legal Research Bench and Tax Agent Bench, and dominates Surge AI’s finance, accounting and spreadsheet modeling tests. For scientific reasoning it tops Terminal‑Bench Science (70.5%), RiemannBench (84.0%) and HLE‑Diamond (71.3%). In coding, Neon is #1 overall on the Artificial Analysis Coding Agent Index (58.6) and posts 65.2% on Terminal‑Bench 4.1.
What are the pricing, token limits and rollout channels?
Neon is priced at $3 per million input tokens, with a heavily discounted $0.15 per million cached input tokens (a 95 % discount), and $15 per million output tokens. The model supports a 1 million‑token context window, enabling very long documents or multimodal collections to be processed without truncation. Availability begins today via the Gemini API in Google AI Studio, the Gemini for business dropdown, and for paid Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app.
How does Neon support production coding and cyber‑defense workflows?
Neon is engineered for long‑running agent workflows that involve many tools and files. It reduces redundant file reads, shares incremental progress, and performs context compaction to stay efficient in extended sessions. In coding benchmarks it leads the Artificial Analysis Coding Agent Index and achieves 46.2% on FrontierCode 2.0 at its default high setting. Multi‑file debugging and targeted repository edits are highlighted as core capabilities.
On the security side, Neon produces more secure code than its predecessor Carbon across the AutoBaxBench suite, reaching a state‑of‑the‑art 76 % on the hard tier. Penetration‑testing performance on WebExploitBench v1.2 is 51.26 % at half the cost of competing large models. Vulnerability‑finding scores 90.5 % on SEC‑bench Pro, while exploitation validation on ExploitGym uses half the tokens of Carbon. Patching benchmarks show a 70.7 % score on CWE‑bench v2, a 3.4 % uplift over Carbon. These capabilities are being rolled out to trusted defenders through the Fairwind Program.
Which enterprise and scientific use cases does Neon target?
Neon’s multimodal context makes it suitable for complex enterprise data that lives in spreadsheets, PDFs, charts, audio and video. The model can turn this heterogeneous input into structured deliverables, a strength reflected in its #1 ranking on Zapier’s AutomationBench (56.3%) and Andon Labs’ Vending‑Bench 2. In scientific domains, Neon excels at quantitative analysis, mathematics and technical research, as shown by its top scores on Terminal‑Bench Science, RiemannBench and HLE‑Diamond.
Showcase demos from Google Antigravity illustrate Neon’s creative reach: a photorealistic 3D lab bench visualizing model names with elemental accuracy, a WebGL2 ski simulator built from scratch with 10,200 lines of custom shaders, an interactive 3D recreation of Alan Turing’s 1936 study, and a transformation of a single image into a 100,000‑cell Google Sheet powered by over 500,000 live formulas.
How can Grafy users experiment with or apply Neon’s strengths?
Grafy creators can explore Neon‑style workflows without leaving the platform. Developers interested in production‑grade coding can test similar large‑context prompts in /code, where the Grafy Code studio lets you run multi‑file debugging sessions and see AI‑generated refactorings. Researchers tackling scientific papers can load PDFs into /reader to query dense mathematical content, mirroring Neon’s performance on benchmarks like RiemannBench. For complex prompt engineering and branching agent logic, the /graph app provides a visual map of prompt variations and model outputs, a useful analogue to Neon’s multi‑step instruction handling. Finally, the multi‑model chat interface in /chat lets you compare Neon‑level responses against other models side‑by‑side, helping you assess cost‑performance trade‑offs before committing to a production API.
Source: https://deepmind.google/blog/introducing-gemini-4-neon/
Frequently asked questions
What token limits does Gemini 4 Neon support?
Neon offers a 1 million‑token context window, allowing it to process very large documents or multimodal collections in a single request without truncation.
How does Neon’s cost compare to other frontier models?
At its default high thinking level Neon delivers a 32 % lower cost per task than Gemini 4 Carbon and roughly 50‑80 % lower cost per task than competing frontier models, while still achieving top benchmark scores.
Can I use Gemini 4 Neon for security testing today?
Neon’s cyber‑defense features are being rolled out to trusted defenders through Google’s Fairwind Program, and the model already shows state‑of‑the‑art results on vulnerability‑finding and penetration‑testing benchmarks.


