You have two vendor quotes open beside a spreadsheet. You need a comparison table, and every price in it needs a page reference so someone else can check it. A quick summary would leave you hunting for those pages yourself.
Start with Hydrogen for a quick edit or summary. Try Gemini 4 Neon when a task regularly sends you back and forth between files or from a code change to its tests. Google announced access through the Gemini app for Pro and Ultra subscribers and through a paid interface for developers.
Which model gets you to a table you can hand off with the fewest trips back to the PDFs?
The right size of job for Neon
Google announced Gemini 4 Neon on November 18, 2026, and began rolling it out that day. Google positions Hydrogen for fast everyday work, Neon for recurring complex analysis and coding, and Carbon for its hardest projects that take longer to complete.
An email edit is an easy starting point for Hydrogen. Use the vendor comparison to try Neon because the answer needs to link facts across files and leave a trail you can follow. For a large refactor, where you change the structure of a codebase while preserving its behavior, put Carbon on your shortlist. If you already use Argon, include it in the same trial. Google describes Argon as a separate model that responds faster and uses fewer tokens per request than Carbon.
Mixed inputs and response size
A Gemini application programming interface, or API, lets your program send requests to the model and receive replies. Through the Gemini API, Neon accepts text, images, video, audio, and PDFs and returns text. Reading across several input formats is called multimodal processing.
A token is a unit of content the model processes, such as a word or part of a word. Neon's context window is the space available for the input and the answer together. The API gives them a shared budget of 1,048,576 tokens, so the files you send take up some of the space available for the reply.
A standard request, where your program waits for the response, can return up to 65,536 tokens. Long decode is a background mode for longer output. It can produce up to 1,048,576 tokens within that same shared budget. For the quote comparison, a short table with traceable cells is a more useful target than a long narrative.
App access and API billing
Google's launch announcement says personal Google AI Pro and Ultra subscribers can choose Neon from the model menu beneath the Gemini app's message box as the rollout reaches them. On the web, open Settings and look for Usage limits to see your remaining allowance. Larger files, longer chats, and settings that let the model spend more effort on a request can use more of that allowance.
Google AI Studio is the browser interface for trying a request before putting it in code. For calls from your own program, enable billing on a Gemini API project and set the model ID to gemini-4-neon. Neon has no free API tier, and you can see the project's active rate limits in AI Studio. A rate limit controls how much traffic your project can send in a given period.
The standard paid API charges USD $3 per million input tokens and $15 per million output tokens. Output billing includes thinking tokens, which the model uses to work through a request before writing the visible answer. When you test Neon from code, record the cost of the request alongside the result so you can compare the full cost of getting a usable answer.
Prompts with a way to check the answer
A prompt is the request you send to the model. Name the inputs, describe the job, and specify how you want the answer delivered. A page number or a test case gives you something concrete to check.
Vendor quotes. “Use the two quote PDFs and the comparison spreadsheet. Put setup fees, monthly charges, and contract terms in one table. Give the filename and page number or spreadsheet row for each value. Write ‘Missing from source’ wherever the files leave a field blank.”
Lecture review. “Read the lecture chart alongside my notes. Explain what each axis measures and find the passage in the notes that explains the pattern. End with two practice questions I can answer from those materials.”
Small code change. “Trace what happens when the input is empty. List the files involved and the order you would change them. Give me a test for empty input and another for valid input, with the expected result for each.” The file list makes the proposed scope easy to inspect before you start changing code.
A side-by-side trial on your files
Zapier ran AutomationBench 1.0.6 on November 17–18, 2026. It gave models 657 tasks in simulated apps through the API, with up to 50 steps per task. At xhigh, the deepest API reasoning setting, Neon passed 370 tasks, or 56.3%. A task passed only when its final state met every condition.
Borrow that last idea for the quote table. Give Neon and your current model the same files and prompt, and make the same tools available to both. Decide up front which cells need a citation and how you will score a blank field. An answer is ready to use when those checks pass.
- Evidence. Open the cited pages and rows and match them to the values in the table.
- Rework. Count the follow-ups you needed to get every required field.
- Time. Record how many minutes elapsed before you had an answer you could use.
- API cost. If you ran the test in code, record what each request cost.
Start with a quote packet you already understand. Check the citations and count the follow-ups first. If Neon consistently saves time on work you often repeat, those savings give you something concrete to weigh against the request cost.
Translated from the Korean original
Sources
- Google’s Gemini 4 Neon launch announcement
- Google AI Studio’s Gemini 4 Neon specifications
- Google Gemini app model guide
- Google Gemini app usage limits
- Google Gemini API rate limits
- Google Gemini API pricing
- Google’s Gemini 4 Argon model description
- Google’s prompt design guide
- Zapier’s Gemini 4 Neon AutomationBench report


Top comments (0)