AI News
AI News AgentModel releaseGoogle4 min read

Google introduces Gemini 4 Neon for complex tasks

Google introduces Gemini 4 Neon, a model for programming, analysis and science with a 1 million-token context window. It comes with four reasoning levels, pricing starting at 3 dollars per million input tokens, and access for developers, enterprises and paid subscribers.

Google has introduced Gemini 4 Neon, a model designed to handle complex programming, analysis and science tasks without the cost or speed tradeoffs of larger models. It is already beginning to roll out to the Gemini API, enterprise products and the Gemini app for paid subscribers.

Neon ranks third overall, with a score of 61, on the Artificial Analysis Intelligence Index when running at its highest reasoning level. According to evaluations cited by Google, it trails larger, slower models but offers a more favorable balance of capability, speed and cost.

A model for long, varied workloads

The model can process up to 1 million tokens, a measure of how much text it can keep in context during a session. In practice, that means it can analyze PDF documents, spreadsheets, charts, audio and video together to generate a report or structured response.

It is also designed for multi-step work. For example, it can review a repository with many files, identify the source of an error, modify several parts of the code and check whether the change works. In a company, the same type of workflow could be used to cross-reference invoices, contracts and financial data before preparing a recommendation.

Google says Neon ranks second on the Vals Index, with 70.9%, and first on Legal Research Bench and Tax Agent Bench. In other tests, it ranks first overall on Zapier's AutomationBench and on Vending-Bench 2, a simulation in which an agent runs a business over an extended period.

These figures come from external evaluations and tests shared by Google, so they do not automatically represent the performance every user will get on a real-world task.

Faster, lower-cost coding

Neon is especially geared toward assisted programming. Google says it ranks first on the Artificial Analysis Coding Agent Index, with a score of 58.6, and achieves 65.2% on Terminal-Bench 4.1.

Its goal is not just to write code snippets. The model can debug applications with multiple files, refactor repositories and work through long sessions using external tools. In Google Antigravity, the company's agent environment, Neon achieves 39% on SWE-sweep, compared with 12.8% on the mini-SWE test used as a reference.

The model has four reasoning levels. You can choose a lower level for fast, inexpensive tasks, or activate xhigh when the work requires more analysis and intermediate steps. At the highest level, Google says the cost per task is 32% lower than Carbon, another model in the Gemini 4 family.

The announced pricing is:

  • 3 dollars per million input tokens.
  • 0.15 dollars per million cached input tokens, with a 95% discount.
  • 15 dollars per million output tokens.

Google also says Neon costs between 50% and 80% less per task than some competing frontier models, although the final savings will depend on conversation length, reasoning level and the tools used.

Security for programming and cyber defense

Google offers a Neon variant focused on cybersecurity. It can help detect vulnerabilities, check whether they are exploitable and propose patches. It scores 90.5% on SEC-bench Pro and reaches 70.7% on CWE-bench v2, according to the company's tests.

The model also includes protections against malicious instructions hidden in web pages, emails, documents or code. This type of attack is known as indirect prompt injection: the agent reads an external source and follows instructions that did not come from the user.

Google says Neon resisted every attack in the Gray Swan benchmark at its highest reasoning level, with an attack success rate of 0%. These are laboratory results, not a guarantee of security for every system connected to the internet.

Where you can try it

Gemini 4 Neon is beginning to roll out across several products:

  • In the Gemini API through Google AI Studio and Android Studio.
  • In Gemini for enterprises, through the API or model selector.
  • In Google Antigravity for Google AI Pro and Ultra subscribers.
  • In the Gemini app for users with Pro and Ultra plans.
  • In the Fairwind program for selected cybersecurity professionals.

Within the Gemini 4 family, Neon sits between Carbon and Hydrogen: Carbon is designed for the longest and most demanding projects, while Hydrogen prioritizes the fast execution of everyday tasks. Neon aims for the middle ground for complex work that needs to be repeated frequently.

The important point is not just that Google has added another model. It is making cost and speed a central part of the choice. For you, that could mean automating more tasks with the same budget, but it also means making a more technical decision: choosing the reasoning level and always checking the results when the model works with code, money or sensitive information.