Gemini 4 Argon's Launch Price Will Double. Which Number Goes in Your YC W27 Application?

On September 30, Google announced Gemini 4 Argon, its new frontier model, and the post went straight to #1 on Hacker News with more than 1,100 points and 750 comments. Most of the thread argued about benchmarks. The detail that matters for YC W27 applicants is in the pricing paragraph and its footnote: Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, and Google states that after the introductory period the price becomes $4 input and $20 output.

That is a scheduled 2x price increase, disclosed on day one, with no end date for the promotion. If your W27 application includes cost math on a frontier model, you now have to decide which number to use. Here is the answer, and how to write it up before the November 2 deadline.

What did Google actually announce?

Gemini 4 Argon was introduced by Koray Kavukcuoglu of Google DeepMind as a model for long-horizon work: software engineering, legal and finance knowledge work, and cybersecurity defense. Google reports 77.9% on DeepSWE v1.1 and raised the output limit to 1M tokens, up from 64K.

Three facts matter more than the benchmarks if you are building a startup on it:

• The price is temporary. $2 input and $10 output is labeled introductory. The footnote sets the standard price at $4 and $20.
• Cached input is 95% off. At the introductory rate that is roughly $0.10 per million cached tokens, which is very cheap for agents that re-read context on every step.
• You cannot use it yet. Argon is rolling out first to "trusted cyber defenders" through Google's Fairwind Program. Developers, enterprises, and consumers get access "as soon as possible," with no date given.

Which price should go in your YC W27 application?

The standard price: $4 input and $20 output. Use the introductory price only as an upside case, and label it as one.

The reasoning is simple. A YC partner reading your application is trying to work out whether the business holds up over years, not over a promotional window. Google has told you in writing what the long-run price is. Cost math built on a number the vendor has already said will double reads the same way as a revenue projection built on a one-time contract.

There is also a neat coincidence in the numbers. As we covered last week when Opus 5.5 and GPT-6 cut prices, GPT-6 Sol is priced at $2 input and $10 output, and Claude Opus 5.5 at $4 and $20. Argon's introductory price matches Sol. Its standard price matches Opus 5.5. So the honest summary for your application is that frontier-quality tokens currently cost between $2 and $4 per million input, and between $10 and $20 per million output, depending on vendor and timing. Plan for the top of that range.

How much does the 2x actually change your margins?

More than most founders expect, because inference is usually the largest variable cost in an AI product.

Take an agent that completes one task using 100K fresh input tokens, 300K cached input tokens, and 30K output tokens, and assume the 95% cache discount carries over to the standard price:

	Introductory price	Standard price
Fresh input (100K)	$0.20	$0.40
Cached input (300K)	$0.03	$0.06
Output (30K)	$0.30	$0.60
Cost per task	$0.53	$1.06

If you charge $2 per task, your gross margin is about 73% at the introductory price and 47% at the standard price. The first number looks like a software company. The second looks like a services business. Same product, same customers, same code. The only thing that changed is a footnote.

Run this table for your own product before you submit. If the business only works in the left column, you have learned something important a month before YC would have asked.

Can you say your startup is "built on Gemini 4 Argon"?

Not truthfully, unless you are in the Fairwind Program. Argon is not generally available, and Google has not committed to a date.

What you can say is that your product is model-agnostic and that you have tested your core workflow on the models you can access today. That is a stronger claim anyway. Three frontier models from three labs have shipped or been announced in roughly nine days. A founder who has wired the product to swap models in an afternoon is describing a real engineering advantage. A founder whose roadmap depends on one unreleased model is describing a dependency.

What if your idea only works at the introductory price?

Then you have four options, and you should name the one you are taking in your application:

1. Route by difficulty. Send the hard 10% of steps to a frontier model and the rest to a small one. Most agent workloads do not need frontier reasoning on every call.
2. Lean on caching. With cached input at 95% off, restructuring prompts so that most of the context is a stable, cacheable prefix can cut cost more than any vendor switch.
3. Move work on-device. On the same day as the Argon post, Magnitude (YC S25) did its Launch HN for a local inference engine for agents, claiming up to 92% faster decoding than llama.cpp on an M4 Pro. Local inference has a fixed cost that no pricing footnote can change.
4. Price on value. If a completed task saves the customer $50 of labor, the difference between $0.53 and $1.06 in cost stops mattering. This is the strongest answer, because it holds at any token price.

How should you write this in the application itself?

Keep it to two or three sentences in the answer about how you will make money. Something with this shape works:

"Each completed task costs us about $1 in inference at standard frontier pricing, and we charge $X. On current promotional pricing our cost is about half that, but we do not plan around it. We route routine steps to smaller models, which brings blended cost to $Y."

That answer does three things at once. It shows you know your cost per unit of work, that you read the pricing page down to the footnote, and that your margins do not depend on a vendor's marketing calendar. Showing all three takes three sentences.

What is the takeaway for W27 applicants?

Last week's story was that frontier prices fell. This week's story is that a headline price can be a promotion. Both are true, and the founders who stand out will be the ones whose numbers survive either direction. YC's Winter 2027 deadline is November 2 at 8pm PT, which leaves about a month to redo the math.

If you want a second opinion on whether your cost and margin claims would hold up when a partner pushes on them, that is the kind of thing YC alumni reviewers on YC Roaster flag when they read an application. Either way, do the two-column table first. It takes ten minutes, and it is a question you should expect to be asked.