Google Unveils Gemini 4 Neon for Development and Data Work

Google has unveiled Gemini 4 Neon, a model for coding, data analysis, and long-running tasks that use tools. The company shared benchmark results, API pricing, and access options.

· 5 min read · 4 comments

Google has unveiled Gemini 4 Neon, a model for coding, data analysis, and long-running tasks that use tools. As Habr reports in its translation of the November 20, 2026 announcement, Google shared benchmark results, API pricing, and access options. Neon is gradually rolling out in the Gemini API through Google AI Studio, in Gemini for business, Google Antigravity, and the Gemini app for Google AI Pro and Ultra subscribers.

What Neon Can Do and How It Performed in Tests

The model has a 1 million-token context window. It can process different types of material—spreadsheets, PDFs, diagrams, audio, and video—within a single context window and produce a structured output. A large context window does not guarantee a correct answer, but it is useful for a team whose information is scattered across files.

Google positions Neon as a model for long-running workflows in which an agent repeatedly uses tools and receives feedback from the user. In Google Antigravity, it can work independently or coordinate Hydrogen subagents in Orchestrator mode. According to the announcement, the model revisits files it has already read less often, reports on task progress, and compresses context during extended work.

In the Artificial Analysis Intelligence Index, Neon scored 61 at the xhigh reasoning level and placed third overall. In the Agent Arena Index, it came second with a score of 1336 across more than 4,200 sessions. These are assessments of behavior under specific conditions, not a promise that every workflow will run without human review.

For development, Google cites a score of 58.6 in the Artificial Analysis Coding Agent Index and 65.2% in Terminal-Bench 4.1. The SWE-sweep result shows the impact of the environment particularly clearly: 12.8% in the general-purpose mini-SWE environment versus 39.0% in Google Antigravity. Comparing only model names without looking at the tools around them would make for a convenient presentation, but would not be very useful when making a choice.

The demos include a WebGL2 freeride skiing simulator built from scratch using elevation data for Mont Blanc and Chamonix. In another example, agents turned an image into a Google Sheet with 100,000 cells, using more than 500,000 formulas without Apps Script or embedded images. These are examples of what is possible, not typical tasks that web teams will pay for every day.

Pricing, Settings, and Access Limits

One million input tokens cost $3 through the API, and one million output tokens cost $15. Cached input tokens are priced at $0.15 per million, a 95% discount. Neon has four reasoning levels: greater depth increases token usage. The default is high; xhigh is intended for the most complex tasks.

Google also compares cost per completed task, not just cost per token. According to the company, at the high level, Neon costs 32% less per task than Carbon. That figure needs to be read alongside the test conditions: usage depends on the reasoning level selected and the number of steps the agent needs.

For security tasks, Google reports results for finding and fixing vulnerabilities: 90.5% in SEC-bench Pro and 70.7% in CWE-bench v2. Google also reports improvements in following instructions and resisting indirect prompt injections. But a separate version, Neon for cyber, is not yet open to everyone: vetted specialists in the Fairwind Program will begin receiving access in the coming days.

What This Means for Web Teams

For developers, the appeal goes beyond code generation to working across multiple files, debugging, refactoring, and long-running tasks that use tools. Site owners and marketers may find value in a workflow that brings information from spreadsheets, documents, and diagrams together into a single analysis. If a team already uses AI in these workflows, it now has another option with published pricing and benchmark results.

I would start a practical evaluation with a specific work task and track three things: the quality of the result, the number of corrections needed, and the total cost. That matters especially for tasks involving access to a repository or internal documents: reported security scores do not replace reviewing changes or a person's authority to decide what goes into production.

My Take

I like that the announcement includes pricing and results across different environments alongside the flashy demos. The gap between 12.8% and 39.0% in SWE-sweep is a reminder of something simple: an agent does not work in a vacuum. The tools and the way the workflow is organized matter just as much as a big name in the model picker.

I would not use a skiing simulator as an argument for buying a model to maintain an ordinary website. But the ability to make sense of a codebase and incorporate feedback as a task progresses is worth testing firsthand. The potential to cut the cost of reused context with cached tokens is also worth checking against your own tasks. Preferably before another experiment quietly becomes part of production.

Sources

Where the news comes from. The text is a retelling in the author’s own words; the facts come from the source, the opinion is the author’s.

  1. [Перевод] Google представляет Gemini 4 Neon habr.com

About the author

Giga Kiknadze

senior full-stack developer

I build web services and maintain several projects using different stacks. I love simple solutions, especially when nobody calls them a “temporary compromise” for five years straight.

All posts by the author

Leave a comment

Not published. We’ll send a link to confirm your comment.

No links or ads. Your first comment is published after review. By sending a comment you agree to the privacy notice.

  1. for web teams, the useful addition is to benchmark tool-call reliability and recovery from malformed outputs, not just coding scores; long-running agents need predictable failure handling before they belong in production

    1. @viktor.hnatyuk in our tooling, the failure that actually hurts is a malformed response that looks plausible enough to slip into the next step. We test retries and validation at each handoff, not just whether the model can write code; a million-token context won’t make a brittle tool chain reliable, no matter how impressive the benchmark chart looks.

    2. One practical test would be whether the agent can recover from a malformed tool response without losing its place in a long task. Before a small team depends on it, I’d want to know how often that recovery needs a person and how much time it adds to routine work.

      1. I ran into this with a small script that chains API calls: one malformed response can throw off every step after it, even when the code itself is fine. I started logging the failed step and retrying only that call, which made recovery easier to inspect; for Neon, I’d want to see whether it can do something similar without quietly repeating work.