Google has introduced Gemini 4 Neon, an AI model aimed at developers and businesses that need complex work completed quickly without always reaching for the largest model. Announced on November 18, 2026, the release brings another option to Google's Gemini family as companies weigh the quality, speed and cost of using AI across coding and analytical tasks.

Neon is available through the paid Gemini API and is rolling out across selected Google products. Google lists a one-million-token context window and API prices of $3 per million input tokens and $15 per million output tokens. Access to the model in the Gemini app and Antigravity is tied to eligible paid subscriptions, so the launch should not be confused with an unlimited free offer. Google's announcement sets out the availability and pricing.
The company is pitching Neon for work that can involve many steps: examining a codebase, reasoning across documents, producing an analysis, and using software tools along the way. Those are claims about the model's intended use, not a promise that it will complete every job correctly. Google's model documentation also acknowledges that it can produce false information.
Where Gemini 4 Neon fits in Google's lineup
Google positions Neon between Gemini 4 Hydrogen, intended for faster everyday work, and Gemini 4 Carbon, aimed at particularly demanding projects. Neon is adapted from Carbon, according to Google DeepMind's model card. The practical question for a developer is whether it can do a particular job well enough, fast enough and cheaply enough to justify using it instead of another model.
That distinction matters because a model selection is not simply a choice between clever and less clever. A customer-support assistant that answers thousands of short questions a day has different needs from a research agent reading a large set of files once a week. The former may be sensitive to the cost of every response and delays experienced by customers; the latter may put more weight on reasoning depth and the quality of citations.
Google describes Neon as suited to software engineering, analytical reasoning and long-running workflows with tools. In a development environment, that might mean reading relevant files, proposing a change and running permitted checks. The model does not acquire independent authority merely because software presents it as an agent. Permissions, tests and human review still determine what it can do safely.
Pricing: what developers would actually pay
At launch, Google lists Neon at $3 for a million input tokens and $15 for a million output tokens. It also quotes $0.15 per million cached input tokens, a 95% reduction from the ordinary input rate for qualifying cached input. Tokens are units used to process text and other material; they are not the same as words, and the number used by a request depends on its contents.
For a simple illustration, imagine an API workload that consumes 100,000 ordinary input tokens and produces 20,000 output tokens. At the published rates, the input portion would cost $0.30 and the output portion another $0.30, for $0.60 in token charges. This example assumes all input is billed at the ordinary rate; it excludes taxes, other service charges and any applicable cached-input pricing. A real bill will depend on the actual usage and terms.
A long context window is not a free allocation of a million tokens. It describes how much context the model can handle, not how many tokens a customer may use without charge. Developers should check the current pricing page, billing settings and usage dashboard before sending large files or enabling automated loops. Repeated tool calls can increase total consumption even when a single answer looks modest.
Comparisons with Carbon also require attention to dates. 9to5Google reported introductory rates for Carbon through November 30 and higher stated rates afterward. A comparison based only on the later rate can mislead readers about what they would pay during the promotion. Prices, availability and eligibility can change, so the provider's current billing information remains the final check.
Is Gemini 4 Neon free?
Not as an unrestricted free API. Google's launch post identifies the paid Gemini API tier for developers, and its app and Antigravity rollout specifies Google AI Pro and Ultra subscribers. Access through a product subscription is not the same thing as free, unlimited API credits. Anyone looking for a no-cost option should inspect the live model list and provider limits rather than assume this particular model is included in a free tier.
Google also says eligible Pro and Ultra subscribers can try Neon in AI Studio. That product access should not be treated as proof that all API usage is free or unlimited. The interface, account region, plan rules and billing arrangement can affect what a user sees. Check the account's displayed terms before starting a large workload.
What a one-million-token context window means
Context is the material the model can consider in a request. A large window can be useful when someone wants to compare several documents, follow a long conversation or examine a sizeable collection of source files. Google's announcement says Neon can work across inputs including documents, spreadsheets, audio and video within its stated window.
Capacity and comprehension are different things. A model may receive a large collection of files and still overlook a crucial detail, misunderstand a table or provide a confident answer that the evidence does not support. Splitting a task into smaller checks, asking for references to source passages and independently verifying important results remain sensible habits.
For sensitive material, there is another practical step before upload: review the service's current data-handling terms and your organization's policies. A large context window can make it convenient to share more information, but convenience does not establish permission to send confidential records or credentials to an external service.
How the coding and agent claims should be read
Google says Neon was designed for long-running, tool-heavy work and describes scenarios such as debugging across multiple files and coordinating smaller agents. The promise is appealing: instead of answering an isolated programming question, an assistant could trace a bug through a project, suggest a patch and run relevant checks in an authorized environment.
The boundaries matter. A suggested patch can pass a narrow test while introducing another problem, and a tool-enabled assistant can misinterpret an instruction or encounter hostile content in a repository. Developers should inspect diffs, restrict access to production systems, use isolated environments where appropriate and run tests that reflect the behavior users depend on. Security-sensitive changes deserve a qualified human review.
Google reports a score of 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at Neon's default high setting. Its model card also lists 65.8% and 47.3%, respectively, for Claude Opus 5.5 in the displayed comparisons. These are useful reference points, but they measure performance on particular test suites under particular setups. They do not establish a universal winner for every language, repository or production task.
The surrounding software matters too. An agent's tools, prompts, access permissions, time limits and test environment can all influence the result. Google's model card reports notably different SWE-sweep scores for different harnesses, a reminder that the model name alone does not explain a complete coding system. Teams should run their own representative evaluations before replacing a workflow.
What the benchmarks show, and what they do not
At its extra-high thinking level, Google says Neon scored 61 on the Artificial Analysis Intelligence Index and ranked third overall at the time of its announcement. Google also reports results for knowledge work, scientific reasoning and agent tasks. These numbers describe snapshots, and rankings can change as new models and evaluation methods arrive.
Benchmarks are most useful when their scope is clear. A score on a curated task set can show how a model performed on those questions; it cannot directly tell an accountant whether a tax calculation is correct, or a software team whether a generated migration is safe to deploy. Internal testing and customer endorsements in a vendor announcement add context, but they are not substitutes for independent evaluation against the reader's own use case.
VentureBeat's coverage also noted a limitation in the speed claims: Google's launch post did not provide absolute completion times for a matched Neon and Carbon workload. That makes it difficult to turn a broad performance description into a reliable estimate of how many seconds a specific job will save.
Who can use it, and where?
Google's announcement lists several routes to Neon, each with its own access conditions. Rollouts and account eligibility can vary, so the safest way to confirm availability is to check the relevant product using your own account.
- Gemini API and Google AI Studio: Google lists the paid API tier and also describes access in AI Studio for eligible Pro and Ultra subscribers.
- Gemini app: Google says AI Pro and Ultra subscribers can select Neon in the model menu.
- Google Antigravity: Google lists access for paid AI Pro and Ultra subscribers.
- Gemini for business: Google says the model is available through its API and model selection interface.
These options do not necessarily have identical limits, features or billing rules. A subscription in one interface should not automatically be assumed to cover a separate API project. Organizations should also check administrative controls and whether their own agreements permit the intended use.
What businesses should test before adopting Neon
Start with a representative task set rather than the most impressive demonstration. For a software team, that could include small bug fixes, a difficult multi-file issue and a change that must preserve an existing interface. For an operations team, it could include summarizing approved documents and extracting values that can be checked against the originals.
Measure more than whether the answer looks polished. Record factual accuracy, the number of interventions needed, task completion time and the total cost of all calls, including retries. For coding, include the time a reviewer spends inspecting and correcting the output. A cheaper response that requires repeated repair may not be cheaper per completed task.
Next, define what the system may read and do. Give it access only to the tools and information required for a particular job, and keep approval steps for consequential actions. Preserve a way to inspect activity and reverse mistakes. These are operational choices made by the deployer, not features a model benchmark can guarantee.
Finally, repeat the test when the model, prompt, tools or pricing change. An apparently small change in how an agent searches files or calls a tool can affect reliability and cost. The goal is not to produce a perfect laboratory score, but to understand whether the workflow is dependable enough for the people who will use it.
Safety, knowledge limits and the missing report
Google DeepMind's model card says Neon can hallucinate and gives a March 2026 knowledge cutoff, with some subjects potentially less current. A current web connection, if available in a particular product, can help retrieve newer material, but the model's summary still needs to be checked against those sources. The cutoff should not be read as a guarantee that every earlier detail is correct.
At the time of the model card consulted for this article, Google said a full Neon safety and security report would be available in the coming days. The card included some automated evaluation results and referred readers to other Gemini safety documentation. Readers should distinguish the materials already available from a full report that had not yet been published in the cited version.
Google also describes cyber-defense uses, including finding and patching vulnerabilities, and says a cyber version is rolling out to trusted defenders through its Fairwind Program. Those announcements do not mean a generated security finding has been validated. A false positive can waste a team's time; a missed flaw can leave users exposed. Authorized testing and independent verification remain necessary.
Why this launch matters
AI buyers are increasingly choosing among models on more than headline intelligence. The economics of an agent that reads files, calls tools and revisits a problem several times can be very different from the economics of a single chatbot answer. If a mid-positioned model performs well on a team's actual work, it may make frequent assistance practical where a larger model is too costly or slow.
That is the opportunity Google is presenting with Neon. Whether it delivers for a particular reader depends on the task, account, safeguards and current pricing. The release is significant because it expands the choices for developers and paid subscribers, not because any model has removed the need to check its work.
Sources and reporting context
This article is based on Google's November 18 announcement and DeepMind model card, alongside reporting by 9to5Google and VentureBeat. Product descriptions, pricing and benchmark figures are attributed to their sources; the practical testing suggestions and cost example are HotPotNews analysis. Availability and prices should be checked again before use.


