Monday afternoon, I dumped three client interviews into one folder—and got stuck

I’m a freelance consultant. On the afternoon of November 18, in a shared office in Nanshan, Shenzhen, I needed to turn three client interviews, website material, and an 80-page industry report into a proposal. The source material was scattered across Feishu, email, and local documents. After 30 minutes of copying and pasting, I was staring at a screen full of text: which passages actually supported my conclusions?

Then I tried Gemini 4 Neon, one of Google’s new Gemini 4 models. It wasn’t a magic answer, but it did compress much of the repetitive work involved in “read the material first, then draft the first version.”

What it is, and who is already using it for real work

Neon is a mid-to-high-end model from Google DeepMind, built to handle very long context with four levels of reasoning effort. Rather than simply filling in a paragraph of copy, I saw it as something that could track relationships and comparisons across an entire pile of material.

Google’s published price is $3 per million input tokens and $15 per million output tokens. Cached input costs $0.15 per million tokens. Since launch, Cursor, GitHub Copilot, and LangChain Managed Deep Agents have integrated it.

My own test happened on Wednesday evening, in a café in Zhujiang New Town, Guangzhou. I put a redacted client interview, public website material, and product documentation into the same chat, asked Neon to list the evidence first, then draft a one-page “client’s current problem—my service recommendation.”

Lin Yue, a friend who consults on financial content, also uses it to read supplier contracts. She has the model identify risks, then checks each one against the original document. It can find leads, but I don’t copy amounts, commitments, or clients’ exact words without verifying them myself.

What it takes to replicate this today

Money: Based on Google’s published rates of $3 per million input tokens and $15 per million output tokens, my mixed task used roughly 100,000 input characters and several thousand output characters. The public list price would come to a little over $0.30; subscriptions and regional pricing may differ.

Time: My first setup took about 60 minutes. Assembling one brief took another 20–30 minutes, plus 15 minutes for manual verification.

Technical barrier: No coding required, but I needed a browser, access to AI Studio or a Google subscription, and enough awareness to redact source material before uploading it.

First step: Sign in with a Google account, open AI Studio, create a new conversation, select Gemini 4 Neon from the right-hand panel, then upload material with no ID numbers, phone numbers, or secrets.

I also got this wrong. My first attempt included 12 documents, and the output was long and scattered—nothing like a clear conclusion. I switched to generating outlines in batches first, then requiring a source reference for every claim. Readability improved dramatically.

Neon may not yet be available in every region or subscription tier, so I wouldn’t rush to pay just to test it. Not trying it now is fine too; I can complete the same workflow with the model already in my toolkit.

How I’d choose it at different stages

If I’m just starting: I’d skip the multi-model experiment and use free allowances to assemble a one-page client brief. First, I’d confirm that the source material can be deleted, properly redacted, and manually reviewed.

If I already have one or two clients: I’d build a fixed template: client background, goals, pain points, evidence, recommendation, and open questions. The goal wouldn’t be to make the model sound more expert, but to make it work in the same structure every time.

If I’m scaling: I’d compare Neon with my existing models on total time for real tasks, not just price per run: uploading, manual verification, and rework included. I’d also establish rules for classifying and retaining client material, so sensitive files don’t casually end up in any model.

If I were deciding whether it belonged in my workflow, I’d run one already-redacted document through it first and see what happened.