The Gemini 4 Argon launch is out for a narrow first wave, with a published token price

Google announced the Gemini 4 Argon launch on 30 Sep 2026. The blog says the model rolls out first to trusted cyber defenders through the Fairwind Program, with broader access later, starting with paid API customers and Google AI Ultra. Reuters, filing from San Francisco the same day, says Google gave no public-release timeline. A flagship post and a missing public clock are both part of the story.

What actually happened

The primary page is Google's own post, Gemini 4 Argon. That is where the rollout order, the prices, the output limit, and the benchmark figures in this pack come from. Reuters, datelined San Francisco on 30 Sep 2026, is the second source.

Reuters adds the size comparison, the missing public timeline, the Gemini 3.5 Pro point, and a spokesperson line. Where one of those facts lives in only one of the two pages, this recap says so.

This item is filed with 6 confirmations and a hype-gas reading of 86. Six fits an announcement that the company blog and a wire story both carry on the same day. Eighty-six is hot.

A new flagship model with benchmark percentages will get screenshotted without the access limits. The limits are the part a reader can actually use. Hot score, narrow door.

Reuters says Argon is larger than Google's previous Pro models. The pack does not give a parameter count, so this article will not invent one. "Larger than previous Pro models" is the comparison Reuters printed.

It is not a leaderboard by itself. Size, in that sentence, is Google's model relative to Google's earlier Pro models, on Reuters' account. It is not a measured claim against every other lab's weights.

Reuters also says Google no longer plans to release Gemini 3.5 Pro. That is a cancellation of a plan, as Reuters reports it, not a review of a model that shipped under that name in this pack. If you were waiting on a 3.5 Pro checkpoint, the news from Reuters is that the plan to release it is off.

Argon is the flagship Google announced instead, on 30 Sep 2026. The pack does not say 3.5 Pro was renamed Argon. Do not glue those names together.

The spokesperson line Reuters carried is a comparison, and the parentheticals are part of how Reuters printed it. The words below are that quote, not the blog's score table.

comparable to frontier models like (OpenAI’s) Astra and (Anthropic’s) Opus on key coding and cyber benchmarks.

On Gemini 4 Argon launch, that is a spokesperson talking to Reuters, not a table reproduced in this pack. "Key coding and cyber benchmarks" is their phrase for which tests they mean. The blog, separately, prints four benchmark figures.

This article will not assume the spokesperson's "key" tests are the same four lines unless the pack says so. It does not say so. Quote the spokesperson as a quote. Quote the blog scores as blog scores.

Who gets it first, and what it costs

Access is a queue, and the blog names the order. First: trusted cyber defenders through the Fairwind Program. Later, broader access, starting with paid API customers and Google AI Ultra.

Reuters says there is no public-release timeline. Put those together and you get a staged opening with an unnamed date for everyone else. "Broader" is not "public."

"Starting with" names the front of the broader wave, not the end of it. Anyone outside paid API customers, Google AI Ultra, and the Fairwind defender set is not given a day in this pack.

The introductory price on the blog is $2 per million input tokens and $10 per million output tokens. Cached input is 95% off. The pack places that cached-input discount with the introductory price, so that is where this recap keeps it.

After the introductory period, the price is $4 per million input tokens and $20 per million output tokens. The pack does not say how long the introductory period lasts. Do not invent a number of weeks. Two price pairs, one unnamed window, one cached-input discount tied to the intro in this telling.

A reader comparing bills needs both pairs in the same note. Intro: $2 in, $10 out, per million tokens, cached input 95% off. Later: $4 in, $20 out, per million.

Doubling is a description of those published pairs, $2 to $4 and $10 to $20, and it is not an extra fee the blog hid. What the blog does not give you here is a cached-input price after the intro. This article will not assume the 95% off survives the introductory period, because the pack attaches it to the intro wording and does not restate it for the later price.

The output limit is a blog figure with a before-and-after inside the same phrase. "1M tokens, up from the previous 64K." One million output tokens, where the previous limit the blog cites was 64K.

That is a cap change on the announcement page. It is not, in this pack, a measured throughput number, a latency number, or a promise that every request can actually spend the full million. The cap is the fact. Operational performance beyond the cap is not in the sources.

Fairwind is a name you should not generalize. The blog says trusted cyber defenders, through that program, are first. It does not list the organizations, the countries, or the acceptance criteria.

"Cyber defenders" is the population named. A general developer with an API key is in the later clause, the paid API customers, not automatically in the first wave. Google AI Ultra is named beside paid API customers as where broader access starts.

On Gemini 4 Argon launch, the pack does not define Ultra's bundle. Do not add features to it.

No public-release timeline, per Reuters, means a date in a reply guy's post is not from these two sources. The announcement day is 30 Sep 2026. The queue is Fairwind first, then broader access starting with paid API and Google AI Ultra.

That is the whole schedule. It is enough to stop a fake countdown and not enough to plan a public launch party.

The benchmark lines on the blog

The blog prints four scores, and this recap copies them as scores on that page, not as an independent rerun. DeepSWE v1.1: 77.9%. AutomationBench: 51.3%.

LVBench: 91.7%. CWE-bench v1: 68%, which the page describes as "ties for first." "Ties for first" is the blog's ranking language on that one benchmark.

The pack does not name who else is tied, and it does not print the other labs' scores. Say 68% and "ties for first." Do not invent the rest of the podium.

These names are the blog's names. DeepSWE v1.1, AutomationBench, LVBench, CWE-bench v1. This article will not explain the test design, the dataset, or the scoring script, because the pack does not.

A high number on LVBench, 91.7%, and a lower number on AutomationBench, 51.3%, are not comparable to each other just because both are percents. Different tests. Different figures. Leave them in their own rows.

The spokesperson's comparison to Astra and Opus sits beside those rows without being stapled to them. Reuters' quote says "key coding and cyber benchmarks." The blog's four lines include names that sound like software-engineering and weakness benchmarks, and this pack does not define them further.

A reader can see the theme. A reader cannot, from this pack alone, audit the spokesperson's sentence against the four percentages. Do not pretend the audit happened.

On Gemini 4 Argon launch, hype gas at 86 is what those percentages will do in a timeline. 91.7 and 77.9 travel faster than the sentence about Fairwind. The reader correction is not to deny the scores.

They are on the announcement page. The correction is to keep the access queue and the missing public timeline in the same post as the scores. A model you cannot call yet is not a model your workflow gained on 30 Sep, unless you are in the group the blog named.

Reuters' "larger than previous Pro models" should not be smuggled into the benchmark list as if size explained 68% or 51.3%. Size is Reuters. Scores are the blog.

The 3.5 Pro plan being dropped is Reuters. The output cap moving from 64K to 1M tokens is the blog. Sort your nouns and the announcement gets clearer, not smaller.

The hype is understandable. The filing is still a staged rollout with a price card.

Confirmation score 6 does not mean every benchmark was replicated by a second lab in this pack. It means the launch facts used here are on Google's post and, where noted, in Reuters the same day. A tie for first on CWE-bench v1 is Google's page saying that.

Treat it as the page's claim. It is specific enough to quote and not specific enough to crown a permanent winner, because no public-release timeline and no named runner-up are in the pack.

What to do as a reader

Separate the door, the bill, and the scoreboard. The door: Fairwind's trusted cyber defenders first, broader access later starting with paid API customers and Google AI Ultra, and no public-release timeline in Reuters' account. The bill: $2 and $10 per million on the intro, cached input 95% off, then $4 and $20 per million after the introductory period, length unnamed. The scoreboard: 77.9% DeepSWE v1.1, 51.3% AutomationBench, 91.7% LVBench, 68% CWE-bench v1 with the blog's "ties for first."

If you are not in that first access set, the practical move is to wait for the broader step the blog described, not to invent a public date Reuters says Google did not give. If you pay API prices, the practical move is to note both price pairs so an intro rate does not become the rate you assumed would last. The pack does not say the intro is permanent. It says there is an introductory period and a later pair of prices.

Use the output-limit line as a cap, not as a workload promise. 1M tokens, up from the previous 64K, is the blog's before and after. It tells you the ceiling the announcement is advertising.

It does not tell you latency, uptime, or which products expose the full window on day one. Those are outside the pack.

Keep the spokesperson on the spokesperson. Comparable to frontier models like OpenAI's Astra and Anthropic's Opus on key coding and cyber benchmarks, as Reuters quoted, parentheticals and all. That is a company voice comparing itself to named models.

It is not the four-row table, and it is not a public download link. Gemini 3.5 Pro, per Reuters, is no longer planned for release. Argon is the model the 30 Sep 2026 posts are actually about.

On Gemini 4 Argon launch, this is not a buying note for any asset, and it is not a claim that a workflow got faster on announcement day. Confirmation score 6 says the launch details above are well sourced. Hype gas 86 says the screenshots will outrun the queue.

Read the queue. Fairwind first. Paid API and Google AI Ultra at the start of broader access. Everyone else's date: not in the release.

Readers who want sourced recaps that are already on the site can read DogeOS public testnet and PUMP token buyback burn as separate live posts.