Google has introduced Gemini 4 Neon, a model positioned between the company’s fastest everyday systems and its more resource-intensive frontier models. The pitch is less about winning a single headline-grabbing test than about handling difficult coding, document and business tasks at a manageable cost. That distinction matters to customers who increasingly pay for the work an AI system completes, rather than for a theoretical ranking.
In its 18 November launch announcement, Google lists an API price of $3 per million input tokens and $15 per million output tokens, with separate pricing for cached input. Those are developer charges, not the price a household pays for a subscription. Google says the model is rolling out through AI Studio, enterprise services, Antigravity and selected paid Gemini subscriptions. Availability differs by product and account.
Why Gemini 4 Neon must be judged by cost per completed task
A business does not buy a model’s abstract intelligence. It buys useful output: a report that can be checked, a software change that works or an administrative task completed without introducing new problems. A model that needs fewer retries can be cheaper overall even if its published price per token is higher. Conversely, a model that excels on a test may still require expensive human correction when faced with messy real-world information.
Google compares Neon with its own Carbon and Hydrogen models. Carbon is intended for particularly demanding tasks; Hydrogen is positioned for faster, lighter work. Neon tries to cover the ground in between. That kind of differentiation allows users to match the difficulty of a task to the resources required, rather than choosing the most expensive model for every request.
Google reports strong scores on coding, spreadsheets and professional evaluations, while third-party rankings vary depending on their design. The company also describes advanced demonstrations involving 3D scenes and substantial spreadsheet calculations. Those examples are useful illustrations of capability, but they are not independent guarantees that a typical user can reproduce the outcome in a different environment.
The importance of the software around the model
One reason AI comparisons are becoming confusing is that the model is only part of the system. A coding agent can inspect files, run tests, call tools and revise errors. Its performance depends on the instructions, available tools, security restrictions and amount of time it is allowed to work. Two products using the same underlying model can produce different results.
Google’s developer announcement for Antigravity describes an orchestration option in which Neon supervises a faster companion model. That does not mean every user automatically receives an autonomous team of agents. The feature belongs to a particular development environment and depends on the tools and access available.
For a manager deciding whether to use AI in a business, the practical test is to compare an actual workflow across models. Give each system the same source files, acceptance criteria and permissions. Record whether the output is correct, how long it takes, how often it needs intervention and what the total process costs. A promotional performance figure may help choose candidates, but this experiment determines whether the product is worthwhile.
What subscribers should check first
Model access varies by plan and region. Google says Neon is available to some Pro and Ultra users and through paid API access, but a product announcement is not confirmation that it has appeared in every account. Users should check the model selector and subscription information rather than assuming that an announced rollout has finished.
The model’s one-million-token API context window is also not a promise that all information supplied in such a long request will be used equally well. Large inputs can help with extensive document sets, but they increase the need for clear task instructions and verification. More context is useful only when the system can identify the right evidence and avoid being distracted by irrelevant material.
LiveAIWire has compared the different strengths of ChatGPT, Gemini and Claude for ordinary users. The release of Neon reinforces the importance of comparing specific jobs rather than declaring one universal winner. A model that is effective at finding a bug may not be the best choice for editing a sensitive letter or evaluating a financial statement.
Powerful models still need boundaries
Google highlights work on security, including resistance to malicious instructions found inside files and websites. These results should be interpreted as evaluations under specified conditions. A benchmark score does not mean every attack will fail, particularly as people change the tools, prompts and data an agent can access.
Where an assistant is permitted to alter company files or send information, security is as important as intelligence. LiveAIWire’s article on AI agents and workplace insider risk explains why organisations need to limit permissions even when they trust a provider’s model. Human review, access controls and reliable logs remain necessary.
The immediate consequence of the launch is more choice for developers and professional users. Whether Neon becomes a preferred model will depend less on its launch-day leaderboard position than on its performance in real tasks where the answer can be independently checked.
The most useful early comparisons will come from people publishing reproducible tests with a clear description of the task, input files, cost and number of retries. A model that looks excellent on difficult demonstrations may be less attractive when a team needs consistent answers to mundane questions every morning. Public benchmarks and provider examples are starting points for evaluation, not substitutes for it. Businesses should also check how changing model versions affects reproducibility before building a permanent process around a newly released system.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.
