Google launches Gemini 4 Neon with top performance for less

Google has launched Gemini 4 Neon, an AI model that handles many of the same demanding tasks as the company's most capable models. In early independent tests, it completed tasks faster than Gemini 4 Carbon and cost less at standard API prices. Google designed Neon for coding, AI agents, advanced analysis, and work that takes several steps. In some tests, it also placed ahead of leading models from OpenAI and Anthropic.
In brief
- Gemini 4 Neon launched on November 18, 2026, and is based on Google's more capable Gemini 4 Carbon.
- Google designed Neon to carry out demanding work faster and at lower cost while retaining much of the capability of its largest models.
- It supports text, images, audio, video, and documents, with a context window of one million tokens.
- Neon scored 61 on the Artificial Analysis Intelligence Index and 70.9 percent on the Vals Index, which includes practical tasks in finance, coding, and law.
- Neon is available through Gemini, Google AI Studio, Google Antigravity, and the Gemini API.
What is Gemini 4 Neon?
Google offers several models in the Gemini 4 family for different kinds of work.
Gemini 4 Hydrogen is designed for fast, everyday work. It suits routine questions, summaries, and simpler tasks where response time matters.
Gemini 4 Neon is designed for more demanding tasks. These include advanced reasoning, coding, analyzing large volumes of information, and completing work through AI agents.
Gemini 4 Carbon is Google's most capable model in this part of the family. It is intended for the hardest tasks, where a solution may take substantial time and computing power.
According to Google DeepMind's model card, Neon was adapted from Carbon. Google has published limited technical details about the adaptation.
The goal is to handle many of the same demanding tasks as Carbon with greater efficiency.
That matters when an AI agent has to work for a long time, use several tools, or process large amounts of data. How fast it works and what it costs matter alongside the quality of the result.
How does Neon compare with Carbon?
The biggest difference is the balance between capability, speed, and cost.
In independent testing by Artificial Analysis, Neon scored 61 on its Intelligence Index at xhigh, its highest reasoning setting. That was one point above Carbon at high.
The scores are close, and the models were tested at different reasoning settings. The gap in task completion time was larger.
In the same tests, Neon took about 21 minutes per task on average, compared with about 36 minutes for Carbon. Neon produced around 110 output tokens per second, compared with about 57 for Carbon.
These averages come from demanding benchmark tasks. A routine chatbot question is much shorter.
Neon used about 15 percent more output tokens per task than Carbon. Its lower prices per token brought the overall cost down.
Artificial Analysis measured an average cost of $6.44 per task for Neon, compared with $7.76 for Carbon at its standard API prices.
In this comparison, Neon finished demanding work faster and cost less despite generating more tokens.
Source: Artificial Analysis, November 18, 2026.
What can Gemini 4 Neon do?
Coding and development with AI agents
Google has made coding a priority for this model.
Neon is designed to understand large codebases, analyze software architecture, find bugs, make changes across multiple files, and test its proposed fixes.
For example, a coding agent asked to diagnose a broken feature in a web service can inspect the relevant files, trace how information moves through the system, make changes, and run tests.
Google highlights Neon's ability to keep track of long sessions, avoid reading the same files repeatedly, and condense the context when a task grows too large to keep all the details available at once.
Neon can also coordinate work in Google Antigravity. It can plan and supervise a task while several faster Hydrogen agents carry out parts of the work in parallel.
With that setup, the more capable model makes decisions and checks the results, while smaller models handle simpler subtasks.
Google's demos show Neon using Blender to build 3D scenes, programming a ski simulator with WebGL, and using multiple agents to build complex spreadsheets.
Those are Google's demonstrations. Consistent performance on comparable tasks would need independent testing.
Analyzing documents, spreadsheets, and financial data
Neon is also designed for what Google calls knowledge work. These are tasks that involve finding, understanding, combining, and using information.
The model can process Excel spreadsheets, Word documents, PowerPoint presentations, PDFs, images, audio, and video. It can use connected tools to retrieve or update information.
Imagine a company that keeps supplier prices in Excel, purchasing records in an accounting system, and contract terms in PDF files.
An AI agent with access to those sources could compare suppliers, flag price changes, investigate discrepancies, and produce a report. With the appropriate integrations, it could also propose or make changes in connected systems.
Neon can analyze information in multiple formats in a single task.
That combination of reasoning, document comprehension, and tool use makes it relevant for automating business workflows.
Science, math, and security
Google also reports strong results on scientific and mathematical problems.
Neon scored 70.5 percent on Terminal-Bench-Science and 84 percent on RiemannBench, two tests covering technical and mathematical tasks.
For cybersecurity, Google describes using the model to identify vulnerabilities in code, assess whether they could be exploited, suggest fixes, and verify the results.
Google is initially making a specialized version, Gemini 4 Neon for cyber, available to approved security organizations through its Fairwind program. Access to that version is restricted to participants.
Four reasoning levels
Neon supports four reasoning levels.
Low: For simpler tasks such as classification, summarization, and extracting structured information.
Medium: For routine coding, document analysis, and tool use when speed and quality both matter.
High: The default, intended for demanding coding, analysis, and work that takes several steps.
Xhigh: The highest level, for especially difficult coding, math, and scientific analysis.
The setting determines how much the model can compute as it reasons before and during a task.
Higher settings can improve results on hard problems, with longer waits and higher token use.
When an AI agent handles hundreds of tasks, choosing a reasoning level appropriate to the work can have a substantial impact on the total cost.
How does Gemini 4 Neon perform in benchmarks?
A benchmark is a standardized test that measures how well an AI model handles a particular kind of work.
Some test problem-solving or math. Others measure how well a model can write code, use tools, or complete an entire task.
Google DeepMind's model card gives the following comparison from November 2026.
| Benchmark | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 61 | 59 | 57 |
| Vals Index | 70.9% | 70.1% | 66.1% |
| AutomationBench | 56.3% | 42.5% | 41.3% |
| Terminal-Bench 4.1 | 65.2% | 65.8% | 61.5% |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% |
| Terminal-Bench-Science | 70.5% | 63.3% | 68.0% |
| Chartography | 73.5% | 66.3% | 75.1% |
| GraphWalks | 95.2% | 66.8% | 80.1% |
Higher scores are better. Reasoning levels and test setups can vary. Google DeepMind's Gemini 4 Neon model card.
Neon is competitive across many of these tests. Claude Opus 5.5 scores slightly higher on Terminal-Bench and FrontierCode, and GPT-6.2 Sol leads this group on Chartography.
The table shows three models. At launch, GPT-6.2 Astra and Claude Fable 5.2 both ranked ahead of Neon on the Artificial Analysis Intelligence Index.
Artificial Analysis tests speed and cost as well as performance
Artificial Analysis is an independent evaluator. It measures model performance, speed, and cost.
Neon's score of 61 on the Intelligence Index at xhigh put it among the top three at launch.
Neon scored about 59 on the Artificial Analysis Coding Agent Index through Google Antigravity. That was about three points above Carbon and close to Claude Opus 5.5 and Claude Fable 5.2 in their respective coding agent environments.
Neon finished the tested tasks faster than Carbon. Smaller models can still complete work more quickly.
The Vals Index tests tasks that resemble professional work
The Vals Index measures how well AI models can handle tasks in coding, finance, law, and tax. It weights results by the economic importance of the relevant industries in the United States.
Neon scored 70.87 percent, placing second behind Claude Fable 5.2 at 71.31 percent.
Neon also cost significantly less to run through this test than several of its closest competitors.
The tests measure how well the models carry out professional tasks of the kind people are paid to do.
Source: Vals Index, version 2.2.
AutomationBench checks whether agents finish the work
Zapier developed AutomationBench to measure how AI agents work across systems for sales, marketing, finance, customer service, and other business functions.
Agents use simulated apps, customer data, email, and other tools. They have to find the information they need and make the correct changes.
The benchmark checks whether the agent completed the requested work in those systems.
Neon scored 56.32 percent at xhigh, taking first place. At high, it scored 54.34 percent at about half the cost per task.
That puts Neon ahead of its competitors in the test. It also shows that even a leading model gets only about half of the tasks completely right under the benchmark's strict criteria.
AI agents still need oversight, especially when working with finances, customer data, or other tasks where mistakes have consequences.
Source: Zapier AutomationBench.
Vending-Bench simulates running a business
Andon Labs developed a more practical test of long-running work.
In Vending-Bench 2, models run a simulated vending machine business for a year. They handle purchasing, pricing, suppliers, and financial decisions.
The goal is to finish the year with the highest cash balance.
Neon came first with an average ending balance of $16,843. Carbon followed close behind with $16,484.
The gap is small relative to the statistical uncertainty. The result leaves open which model performs better on this test.
Source: Andon Labs, Vending-Bench 2.
How much information can the model handle?
Gemini 4 Neon has a context window of about one million tokens.
A token is a small unit of text the model processes. A word can take up one or more tokens.
A context window this size gives the model room for long documents, large parts of a codebase, or information from several sources at once.
Google's developer guide describes a shared limit for input, internal reasoning, tool calls, and output. A single response defaults to a limit of 65,536 output tokens, and developers can use continuation to generate longer responses.
A large context window gives the model space for more information. Tests such as GraphWalks measure how well the model can actually find and follow details across that context.
Neon's 95.2 percent score on GraphWalks is a positive result for work with large codebases and long documents.
How much does Gemini 4 Neon cost?
Google lists the following API prices.
| Type | Price per million tokens |
|---|---|
| Input | $3.00 |
| Cached input | $0.15 |
| Output | $15.00 |
Cached input is information from an earlier request that the model can reuse. Google discounts those input tokens by 95 percent.
Neon costs 25 percent less per token than Carbon's standard API rates of $4 for input and $20 for output.
Carbon's introductory rates of $2 for input and $10 for output run through November 30, 2026. That makes Neon more expensive per token during the promotion. Neon will have the lower rates once the promotion ends.
The total cost of a task also depends on how many tokens each model uses.
In the Gemini app, usage is governed by the subscription and its limits. The API prices above apply to developers using the API.
Where is Gemini 4 Neon available?
Google started rolling out Neon on November 18.
The model is becoming available in the Gemini app for Google AI Pro and Ultra subscribers. Google also offers it in Google AI Studio, through the Gemini API, and in Google Antigravity for eligible subscribers.
The model ID for developers is gemini-4-neon.
Vercel also offers the model through AI Gateway using the model ID google/gemini-4-neon. Developers can integrate Neon into their own apps and AI systems.
Access, quotas, and pricing depend on the platform.
Limitations and safety
Google lists a knowledge cutoff of March 2026 for Gemini 4 Neon. The model may need up-to-date sources such as web search or connected systems for later events.
It can still hallucinate, producing plausible information that turns out to be wrong.
Google reports fewer cases of the model taking shortcuts, claiming to have completed work it did not do, or acting outside its instructions.
In one internal test measuring whether an AI agent accurately reports its actions, Neon scored 99.2 percent.
These are internal results. Safety in real systems would need independent evaluation. Google's full safety report was still forthcoming at launch.
Businesses giving AI agents access to financial systems, databases, or external systems still need access controls, logging, validation, and human approval where mistakes could have serious consequences.
What to take away
Gemini 4 Neon is designed to make demanding AI work more practical and affordable. In launch tests, it competed with some of the best models and completed tasks faster than Carbon at a lower cost using standard prices.
Its strongest results include AI agents, business tasks, coding, and analysis of large amounts of information. The benchmarks also show that even leading agents make frequent mistakes during complicated work they have to complete on their own.
Developers and businesses can choose a model based on the task. Smaller models can handle routine work, Neon can take on demanding work that comes up often, and the most capable models can be reserved for the hardest tasks.
The useful comparison is how reliably a model completes the work, how long it takes, and what it costs.
Sources
- Google – Introducing Gemini 4 Neon, November 18, 2026
- Google DeepMind – Gemini 4 Neon Model Card
- Google Cloud – Developer's guide to Gemini 4 Neon
- Artificial Analysis – Independent evaluation of Gemini 4 Neon
- Vals AI – Vals Index
- Zapier – AutomationBench
- Andon Labs – Vending-Bench 2
- Google – Gemini 4 Carbon and API pricing
- Vercel – Gemini 4 Neon on AI Gateway