Daily · Article 02 · 4 min read
Gemini 4 Neon in the app and API, with a worked bill
November 19, 06:00 JST · announcement dated November 18
Summary
Google is rolling Neon out to Gemini Pro and Ultra subscribers, with general release on its paid API. A worked example shows the shared token budget and Standard charges.
Key facts
Google introduced Gemini 4 Neon on November 18, the date on its announcement. The model can read material such as documents and charts and return text for tasks that span several steps. Google is rolling it out to Pro and Ultra subscribers in the Gemini app. The separate Gemini API, which lets software request the model directly, lists gemini-4-neon as generally available on its paid tier. Google announcement · API release notes

1. One task, two ways to request it
Suppose a shop wants to compare two suppliers' price sheets. An owner can open the Gemini app, attach available material and ask which rows require a follow-up. A developer can use the API to send the same kind of request whenever a new list arrives, then put the result in a review queue. In both cases, compare any quoted price with the original row, especially when the sheets use different package sizes.
| Route | Who uses it and what they can do | Access and billing | Source |
|---|---|---|---|
| Gemini app | A person asks the model to read and compare price sheets | Rolling out to Google AI Pro and Ultra. Check the signed-in account's model picker | Rollout announcement |
| Gemini API | Software sends the sheets and fields to extract, then receives text | Generally available on the paid API tier. Check project billing and access to gemini-4-neon | API model · Rates |
- Route
Gemini app
- Who uses it and what they can do
- A person asks the model to read and compare price sheets
- Access and billing
- Rolling out to Google AI Pro and Ultra. Check the signed-in account's model picker
- Source
- Rollout announcement
- Route
Gemini API
- Who uses it and what they can do
- Software sends the sheets and fields to extract, then receives text
- Access and billing
- Generally available on the paid API tier. Check project billing and access to
gemini-4-neon
The cat asks Q&A
- The researcher
The API lists Neon. Can I select it in my own Gemini app today?
Pocket NotebookCheck the model picker while signed into your Pro or Ultra account. The API's release status tells you about a different route, with billing attached to the developer's project.
2. A long reading budget still has to fit the reply
A token is a piece of text or a unit the service uses to count other inputs. A context window is the model's working space for one request. The API model page gives Neon a shared 1,048,576-token input-and-output window. The default synchronous reply, where the caller waits for an answer to this request, has a maximum output of 65,536 tokens. A separate long-decode mode allows up to 1,048,576 output tokens within the shared limit.
For example, imagine the service counts all input for a request at 1,000,000 tokens. Subtraction leaves 48,576 tokens in the shared window, so the default synchronous limit of 65,536 does not give that request 65,536 tokens of room. For a supplier comparison, start with the relevant sheets and request the source row for each finding. A larger input allowance alone does not guarantee that the model cites each figure correctly. Model limits
The cat asks Q&A
- The researcher
Can I put in a million tokens and get another million back in the same call?
Pocket NotebookThe input and output draw from one shared window. Allow room for the response before sending the documents.
3. Calculate the bill with the right token categories
In Standard mode, Google charges US$3 per million ordinary input tokens, US$0.15 per million cached input tokens, and US$15 per million output tokens, including thinking tokens. Cached input is repeated material the service successfully reuses. Neon supports automatic caching rather than explicit cache creation. In this illustration, the assumed cache hit is a measured category in the service's usage report, not a promise that resending a document gets it. Standard pricing
| Fictional API request | How the Standard charge is calculated | Total |
|---|---|---|
| 80,000 ordinary input tokens and 4,000 billable output tokens | 80,000 ÷ 1,000,000 × US$3 + 4,000 ÷ 1,000,000 × US$15 | US$0.300 |
| Of the same 80,000 input tokens, 60,000 are reported as cached | 20,000 ÷ 1,000,000 × US$3 + 60,000 ÷ 1,000,000 × US$0.15 + the same US$0.060 output | US$0.129 |
- Fictional API request
80,000 ordinary input tokens and 4,000 billable output tokens
- How the Standard charge is calculated
- 80,000 ÷ 1,000,000 × US$3 + 4,000 ÷ 1,000,000 × US$15
- Total
- US$0.300
- Fictional API request
Of the same 80,000 input tokens, 60,000 are reported as cached
- How the Standard charge is calculated
- 20,000 ÷ 1,000,000 × US$3 + 60,000 ÷ 1,000,000 × US$0.15 + the same US$0.060 output
- Total
- US$0.129
This is a calculation, with no search, other tool charges or taxes. Holding output fixed isolates the cache effect on this example's total. The 95% discount applies to the cached input rate. The whole example bill falls 57%.
The cat asks Q&A
- The researcher
Should I budget for every request at the discounted rate?
Pocket NotebookEstimate from the ordinary input and billable output first. Recalculate once the usage report shows how much input was actually cached.
To close
Compare one small task before changing the workflow
In the app, check model availability and compare a few cited source rows. With the API, use the same sample to check output, billed tokens and the cases that still require human review.
Sources
- Rollout announcementblog.google
- API release notesaistudio.google.com
- Model limitsaistudio.google.com
- Standard pricingaistudio.google.com