Junious Media NetworkSaturday, November 21, 2026
Juni/Tech
Advertisement
AI

Google Launches Gemini 4 Neon for AI Pro Subscribers

Google rolled out Gemini 4 Neon, a model built for complex coding and analytics that slots between Hydrogen and Carbon in the Gemini 4 lineup.

Key points

  1. Google rolled out Gemini 4 Neon for AI Pro and Ultra subscribers, placing it between Hydrogen and Carbon in the Gemini 4 family.
  2. The model is priced at $3 per million input tokens and $15 per million output tokens, with a 1 million-token context window and four thinking levels.
  3. Neon replaces Argon and Gemini 3.1 Pro in the model lineup, 9to5Google reported.
  4. On the Artificial Analysis Intelligence Index, Neon placed third overall, with only larger and slower models ranking above it.

Google rolled out Gemini 4 Neon, expanding its Gemini 4 family with a model aimed at demanding professional work in software engineering, data analysis, and science. The new model sits above the speed-focused Hydrogen and below the heavy-duty Carbon in that lineup, targeting workflows where deep reasoning matters but where a slower model would cause delays, Google said.

Following the previous week's release of Carbon and Hydrogen, Google has now filled in the middle tier of its Gemini 4 lineup. Neon replaces Argon and Gemini 3.1 Pro in the app's model selection menu, 9to5Google reported. Gemini 3.5 Flash-Lite became the automatic choice for paid users at the same time, 9to5Google reported.

Where Gemini 4 Neon Fits the Lineup

Carbon handles the longest and hardest projects. Hydrogen covers fast, routine tasks. Neon fills the space between, handling high-frequency workflows that call for deep reasoning at speeds closer to Hydrogen than Carbon.

The model is now available to AI Pro and Ultra subscribers in the Gemini app, as well as in Google Antigravity and the Gemini API, Google said. Enterprise customers can reach it through the Gemini for business API or the model drop-down inside Antigravity. That covers individual users, developers, and corporate teams through whichever platform fits their workflow.

Pricing and Context Window

Google set the API price at $3 for every million input tokens and $15 for every million output tokens, Google said. Teams that use caching pay $0.15 per million input tokens, a 95% reduction from the standard rate. That puts Neon at twice the cost of Hydrogen on both input and output tokens, 9to5Google reported.

Advertisement

Argon and Carbon currently sit at a promotional rate of $2 per million input tokens and $10 per million output tokens through November 30, after which those prices will climb to $4 input and $20 output, 9to5Google reported. After that change, Neon will cost about 25% less than Carbon, 9to5Google reported. At its default High thinking level, Neon delivers a 32% reduction in cost per task compared to Carbon, Google DeepMind said.

Advertisement

The model comes with a 1 million-token context window and four thinking levels: Low, Medium, High, and Extra High, 9to5Google reported. That range lets teams set the model's reasoning depth to match each task. Lighter jobs can run at lower settings, while complex analytical or coding projects can draw on the deeper levels for greater precision.

Benchmark Performance

Neon placed third overall on the Artificial Analysis Intelligence Index, posting a score of 61 at its top thinking level, with only larger and slower models ranking above it, Google said. It scored second on the Vals Index at 70.9%, and took the top spot on both the Legal Research Bench and Tax Agent Bench. For reading dense professional charts, the model scored 73.5% on Chartography, Google said.

Science and math results were broad. Neon placed first on Terminal-Bench Science at 70.5%, on RiemannBench at 84.0%, and on HLE-Diamond at 71.3%, Google said. On Agents' Last Exam v1.1, its score of 75.8% matched the results of leading models. Those evaluations cover quantitative analysis, mathematics, and technical research problems, making them a wide measure of scientific reasoning.

Advertisement

Coding benchmarks also showed strong results. Neon ranked first on the Artificial Analysis Coding Agent Index at 58.6 and hit 39.0% on SWE-sweep inside Google Antigravity, Google said. It also topped Zapier's AutomationBench at 56.3% and won Andon Labs' Vending-Bench 2, a long-horizon business simulation, with a result of $16,843.

Built for Long Agent Workflows

Neon is designed to stay consistent through extended, tool-heavy sessions without drifting off-task, Google said. It can work on its own or as an orchestrator, managing parallel Hydrogen subagents through multi-step pipelines inside Google Antigravity. The model is built to cut redundant work, share progress as tasks run, and handle context reliably across long sessions.

On the Agent Arena Index, Neon scored 1336 across more than 4,200 real-world agent sessions, ranking second globally, Google DeepMind said. That index includes a steerability measure, tracking how well a model responds when a user changes direction mid-task. Neon ranked second on that measure, which carries weight in automated workflows where requirements shift during execution.

Cybersecurity Capabilities

Google built a security-focused version called Neon for cyber, which generates more secure code than Carbon across all difficulty tiers on the AutoBaxBench evaluation, reaching 76% on the hardest challenges, Google said. On WebExploitBench v1.2, a penetration testing benchmark, Neon scored 51.26% at half the cost of comparable large models. It also achieved 90.5% on SEC-bench Pro for finding vulnerabilities in existing code.

Advertisement

Neon for cyber will reach qualified members of the Fairwind Program over the next several days, the company said. That program serves cybersecurity defenders who use AI tools for vulnerability analysis, patch validation, and exploitation testing. Enterprise teams seeking access can contact their Google Cloud account representative.

Advertisement

Safety and Model Behavior

Google said Neon posted a 99.2% pass rate on its internal agent honesty evaluation and a 0.09% misaligned action rate on the company's agentic safety suite. It matched Carbon with a 0.0% cheating rate on the Terminal-Bench 4.0 evaluation and recorded the lowest shortcut rate of any Gemini 4 model at 2.45% on the coding cheating test, Google said. Those results suggest the model is less likely to cut corners or fabricate results when handling long, complex tasks.

On Gray Swan's IPI benchmark, which tests whether a model can be manipulated through malicious instructions buried in documents, emails, or code it reads, Neon stopped every planted attack when its highest-level real-time safeguards were active, Google DeepMind said. Among models in its class, only other Gemini models have posted a 0.0% attack success rate on that test, Google DeepMind said. For enterprise teams running automated agents that process outside content, that resilience has practical value.

What It Means for Subscribers

Consumers who pay for Google AI Pro or Ultra can select Neon from the model drop-down in the Gemini app. For anyone who regularly uses AI for tasks like analyzing documents, drafting complex reports, or reviewing code, Neon offers more reasoning depth than Hydrogen without the added time and cost of Carbon. It also reads uploads such as spreadsheets, PDFs, and photos and can pull in context from connected Google apps like Gmail and Docs, Google said.

Businesses running automated pipelines or long agent workflows can access Neon through the Gemini for business API or inside Google Antigravity, Google said. The Gemini 4 family now spans from Hydrogen's quick everyday tasks up to Carbon's most demanding projects, with Neon covering the complex high-frequency work between them.

Advertisement

Sources