Gemini 4 Argon is not a public product launch. On September 30, 2026, Google posted a frontier model, a token price, and a narrow gate. Koray Kavukcuoglu, senior vice president of Google DeepMind and the company’s chief AI architect, wrote that the model was rolling out to a set of trusted cyber defenders through the Fairwind Program. Developers, enterprises, and consumers were told access would come “as soon as possible,” after guardrails, not on the day of the post. That is the reverse of the ChatGPT pattern, in which a consumer URL and an API key arrived in the same news cycle. Alphabet still gave traders a number. Nasdaq’s regular-session close for Class A shares, ticker GOOGL, was $344.08 on announcement day. By Friday, October 2, GOOGL finished at $343.50, up 1.56% on the session, and Nasdaq’s quote summary listed a market capitalization of $4.201 trillion. The rate card is public. The meter is not.
The comparison that fits is an exchange opening a new matching engine. Designated market makers touch the system first. Everyone else gets a circular, not a live order ticket. Gemini 4 Argon is that sequence with a cloud price list attached. A buyer can already model a bill: $2 per million input tokens and $10 per million output tokens, cached input at 95% off, then $4 and $20 when the introduction ends. The same buyer cannot send production traffic. Several write-ups called the technical jump a 16-fold context window. Google’s post says something narrower. The output-token limit rises from 64,000 to 1 million, a ratio of 15.625 times. That is about 16-fold only as a cap on how long the model may keep writing, not as a published context-window spec. I would not put Argon consumption into a Google Cloud revenue model for the current quarter. The quarter can price the option. It cannot invoice it.
Key facts
- On September 30, 2026, Google said Gemini 4 Argon was going first to trusted cyber defenders through the Fairwind Program, with developers, enterprises, and consumers to follow “as soon as possible.” Source: Koray Kavukcuoglu’s post on the Google blog.
- The introductory price is $2 per million input tokens and $10 per million output tokens. Cached input is 95% off the input price. After the introduction, Google’s footnote sets $4 and $20. Source: Google.
- Google said the output-token limit is 1 million, up from 64,000, or 15.625 times the prior cap. The post does not call that change a context window. Source: Google.
- Nasdaq’s historical series shows a September 30 close of $344.08 and an October 2 close of $343.50, up 1.56% that Friday. The quote summary listed a market cap of $4.201 trillion. Source: Nasdaq.
- Investing.com reported that Alphabet shares rose about 2% after hours on September 30. The next session opened at $350.785 and closed at $338.24. Sources: Investing.com via Yahoo and Nasdaq.
- Sundar Pichai signed a voluntary White House AI safety accord on September 29, the day before the model post, alongside leaders from Anthropic, OpenAI, Meta, xAI, and Nvidia. Sources: Al Jazeera and CNBC.
- Google claims a DeepSWE v1.1 score of 77.9% and a tie for first on CWE-bench v1 at 68%. Those are company figures, not an independent audit. Source: Google.
What Gemini 4 Argon is
Google describes Gemini 4 Argon as a model for jobs that do not finish in one prompt. Kavukcuoglu’s September 30 post says it targets software engineering, knowledge work such as legal and finance, and cybersecurity defense. Google says thousands of employees are already using it for coding, research, and writing. The examples it chose to publish are infrastructure jobs, not a consumer demo.
Google says one quantum run beat a published baseline on qubit-and-gate resources by 40% within minutes, and that memory agents freed more than 300 tebibytes, with an estimate of 500 tebibytes to one pebibyte still ahead. It says code agents are moving C and C++ toward Rust, up to more than 800,000 lines in the Fuchsia Zircon kernel, still in audit before production. For libgav1, it says the new decoder was 2.7 times faster than an earlier Rust port, with the same video output. Those are company figures, not an outside re-test.
Google says Argon is state of the art on DeepSWE v1.1 at 77.9%, first on Zapier’s AutomationBench at 51.3%, first on LVBench at 91.7%, and tied for first on CWE-bench v1 at 68%. It also says Argon leads the Vals Index. Rival scores that are not in the post are omitted. The output limit, 1 million tokens against a prior 64,000, is what changes a bill. A long run can stay in one generation, and output tokens are the expensive side of the card.
The live list is Fairwind, plus Google’s own teams. Google says those defenders receive the model without cyber guardrails, and it names Wiz, through Scan for Good, as an early user. Google says a demonstration found a critical flaw in healthcare software that earlier frontier models had missed. It does not name the flaw. Before a wider release, the post lists misuse refusals, covering cyber requests and chemical, biological, radiological, and nuclear requests, plus stronger prompt-injection defense, monitors on reasoning and actions, and tighter sandboxes. Commercial access, Google says, starts with paid API customers and Google AI Ultra subscribers, not with a free switch in the Gemini app.
Tulsee Doshi, product lead for Gemini, told CNBC why a price can be public while a signup page is not. “Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible,” she said.
What rivals and buyers have actually said
In the reports read for this piece, Microsoft and Amazon had not issued an on-the-record reply to Gemini 4 Argon by Sunday, October 4. The record instead holds a price comparison, a safety hold at OpenAI, two sell-side notes, and a consumer-agent contest that is a different product.
CNBC reported on October 1 that the introductory $2 and $10 rates match OpenAI’s newly discounted GPT-6.1 Sol. That pairs a discounted OpenAI card with a Google card that is published and still closed. It should not be merged with a second report. Al Jazeera wrote on September 29 that OpenAI would not release a model it called GPT-6.1 Astra after in-house safety testing. The names are close. They are not the same claim, and folding them into one SKU would invent a product neither outlet described.
Anthropic did not answer with a Claude scorecard. CNBC noted that Dario Amodei had been urging labs to slow development, and on September 29 he signed the White House accord beside Sundar Pichai. That is a regulatory fact next to a product launch, not a benchmark rebuttal.
JPMorgan analysts wrote on Thursday that Google still needs a step forward in personal agents before consumers care. Bank of America, in the same CNBC report, pointed at Google Cloud, at products Google already runs, and at a later agent. A model a client cannot call does not enter cloud backlog. Spark, Google’s agent, stays behind a paywall, while Meta’s Muse is free with caps. CNBC, citing Sensor Tower, said Muse had passed 5 million downloads as of September 30, a race FinanceFeeds has followed beside Alphabet’s capital-spending plans and Meta’s Muse. Doshi told CNBC that cheaper models still handle ordinary agent traffic, and that frontier weight is for long jobs.
Outside that cyber list, a buyer can read the card and cannot place an order. Wiz, the outside user Google named, is a defensive-scanning partner, not a general reference. CNBC relayed a Bloomberg report of employee doubts about everyday coding, and Google’s reply that staff had tested Gemini 4 for weeks. Outsiders still cannot run that test.
The sharpest line on the launch as a launch was the JPMorgan sentence CNBC quoted. “Google has struggled to translate ongoing product velocity into splashy releases that resonate with consumers, capture sustained mindshare, & meaningfully drive sentiment around AI leadership,” the analysts wrote.
The stock, the rate card, and the math
September 30 was not a clean gap. GOOGL opened at $344.15, hit $352.60, and closed at $344.08, up 0.93% from $340.92, on 37.4 million shares. Investing.com put the after-hours move at about 2%. October 1 opened at $350.785, 1.9% above that close, ranged from $353.22 to $335.51, and finished at $338.24, down 1.7%. Friday’s close was $343.50, up $5.26, or 1.56%, on 23.8 million shares. With no Saturday session, that is the last official close.
From the July 1 close of $361.21 to October 2 at $343.50, Class A is down 4.9%. The low close in that span was $317.69 on July 23, and the high close was $377.65 on August 4. Friday sits 15.9% under the 52-week high of $408.61 and 45.6% above the 52-week low of $235.84. Nasdaq’s summary lists a $430 one-year target, 25.2% above $343.50. FinanceFeeds has published a separate scenario on Class C shares after the same announcement. Class C is not the GOOGL series used here.
Five million input tokens and one million output tokens cost $20 at the introduction, five times $2 plus $10, and $40 after the footnote, five times $4 plus $20. Both rates double, so the bill doubles. If every input token is a cache hit, and the 95% discount rides on the input price then in force, input falls to $0.50 and later $1.00. With output, the totals are $10.50 and then $21.00. The generated tokens, not the cache, dominate. FinanceFeeds has already laid out the $2 and $10 introductory card. Google did not date when that introduction ends.
| Question | Bull reading | Bear reading |
|---|---|---|
| Token price | $2 and $10, cached input 95% off, is modelable. CNBC says it matches discounted GPT-6.1 Sol. | The footnote resets the card to $4 and $20. If that hits before access, $2 was never the invoice. |
| Access | Fairwind and the government pre-release give Google a controlled reference set before a wide opening. | No public meter means no Gemini 4 Argon consumption line in Google Cloud this quarter. |
| The stock | Friday’s close was $343.50, up 1.56%, with Nasdaq’s market cap at $4.201 trillion and a listed one-year target of $430. | October 1 opened at $350.785 and closed at $338.24. From July 1 the shares are down 4.9%, and 15.9% below the 52-week high. |
| Evidence | Claimed scores, including 77.9% on DeepSWE v1.1 and a 68% tie on CWE-bench v1, are specific enough to test later. | They are company scores. Employee skepticism reported by CNBC, which Google disputes, stays unsettled until outsiders can run the model. |
| Cloud contest | Once keys exist, cache economics and a 1 million output cap are the wedge versus a rival already at $2 and $10. | Until then, another vendor’s live model wins a production bake-off. A blog post is not a security review. |
The accord, the pre-release, and what is not a ban
The model post landed one day after a different Google signature. On Tuesday, September 29, President Trump hosted AI executives at the White House. CNBC reported that Pichai signed a voluntary safety accord with the president and other technology leaders, and that the Argon announcement came a day later. Al Jazeera named six signatories: Anthropic’s Dario Amodei, OpenAI’s Greg Brockman, Google’s Sundar Pichai, Meta’s Mark Zuckerberg, xAI’s Elon Musk, and Nvidia’s Jensen Huang. SiliconANGLE’s list is the same six.
The title is not stable across those accounts. Al Jazeera called the text the Joint Commitment on Frontier Responsibilities. SiliconANGLE called the document Trump posted on Truth Social late Tuesday the White House Accord on Super Intelligence. The White House file itself was not among the sources read here, so neither headline is treated as a confirmed short title. Both outlets describe a voluntary pledge: internal controls, outside auditors, and regular meetings on safety practices. Neither describes a statute with a fine schedule.
Two quotations show the limit of the promise. Al Jazeera quoted Trump at the luncheon: “There’s a belief that there should be tremendous self-regulation, and we automatically have regulation with the Department of Justice, the FBI, all of that,” he said. SiliconANGLE quoted the signatories: “We believe every company is responsible for developing its own technology safely and in a way that builds trust with customers and the public,” the group stated. Al Jazeera added that critics focused on the nonbinding character of the text. A Tuesday pledge does not set Wednesday’s token price.
The Argon post describes a second gate: the U.S. government’s voluntary process for pre-release model access. That is not the chief-executive accord. One is a public pledge by named executives. The other is how Google says this model is shown before a broad release. No Argon-specific export order appeared in the Google post, the CNBC reports, or the accord coverage used here. A closed API should not be pinned on a Commerce Department hold that those pages do not mention.
Fairwind and Google’s own teams are inside. Paid API customers, AI Ultra subscribers, and consumers are not, until misuse refusals, prompt-injection defenses, reasoning monitors, and sealed sandboxes move. That list is what a regulated buyer will ask about later. The accord may color the review. It does not open the meter.
What has to happen before the price means anything
Three forecasts follow from causes already on the page. None of them is a ship date, because Google did not give one.
The next commercial step, if Google follows its own sequence, is a paid API and Google AI Ultra, not a free button in the consumer app. The cause is the order in the September 30 post: feedback from the cyber cohort, iteration on guardrails, and the voluntary pre-release, then developers, enterprises, and consumers, starting with those paid doors. A same-week public chatbot would contradict that order. The falsifier is a developer note that turns the $2 card into a live SKU. Until that note exists, “availability” still means a gate.
The post-introduction price is already published, at $4 per million input tokens and $20 per million output tokens. The unknown is whether anyone outside Fairwind meets that card or the $2 card. If the pre-release ends while the introduction is open, Google Cloud can enter a bake-off at the rate CNBC tied to GPT-6.1 Sol, plus cache discounts and a longer output allowance. If the introduction expires first, $2 was a press-cycle number. The cause is the missing expiry date. Procurement should price both the $20 and $40 examples above, and should not assume the lower bill survives into a contract.
Google Cloud cannot book consumption on a model customers cannot call, and neither Azure nor AWS has had to answer in public to keep today’s work. No price cut from Microsoft or Amazon showed up in the reporting checked here, so none is imputed. A frontier meter needs a key. Memory savings Google claims are an Alphabet cost, not a Cloud invoice. The cash strain is already the subject of Alphabet’s free-cash-flow pressure from AI spending, and the silicon is covered in Google’s in-house accelerator work. Those bills are due whether or not a client can buy Argon this month. Once a key exists, beating a $2 rate CNBC says is already matched will not settle the contest. Reused cache, output room for long jobs, and distribution inside current Google Cloud accounts might.
FAQ
Can anyone outside Google’s cyber list buy Gemini 4 Argon today?
No. The September 30 post limits the current rollout to trusted cyber defenders in the Fairwind Program, while Google says it is also in a voluntary U.S. pre-release process. Developers, enterprises, and consumers were promised access “as soon as possible,” starting with paid API customers and Google AI Ultra subscribers. No date was published. A token price and a working API key are different events, and only the price has happened.
What does Gemini 4 Argon cost after the introductory period?
Google’s footnote sets $4 per million input tokens and $20 per million output tokens, double the introductory $2 and $10. Cached input is described as 95% off the input price. Google did not say how many days the introduction lasts, or whether the cache discount is rewritten when the list price changes. Until those terms are dated, a buyer cannot know if $2 will still be offered when access arrives.
Did Alphabet stock rise 2% because of the announcement?
Investing.com reported an after-hours rise of about 2% on September 30. The Nasdaq close that day was $344.08, up 0.93% from $340.92, after a high of $352.60. On October 1 the shares opened at $350.785 and closed at $338.24. Friday’s close was $343.50, up 1.56%. The 2% figure is the evening tape, not the regular-session close, and the next session gave that gap back.
What is the Fairwind Program?
Fairwind is Google’s name for the first handoff of Gemini 4 Argon to trusted cyber defenders. The company said those defenders, and its own teams, would receive the model without cyber guardrails so they could use its full defensive capability. Other customers wait while safeguards are revised. Google published no member list. It did say Wiz is using the model in Scan for Good, including on an unnamed healthcare flaw.
Does the White House accord stop a wider Gemini 4 Argon release?
Not on the reporting in hand. The September 29 accord is voluntary. Al Jazeera and SiliconANGLE describe internal controls, outside auditors, and standards meetings, and Al Jazeera’s critics stressed that it is not a binding statute. Google separately cites a U.S. voluntary pre-release process. That process, plus the guardrails in the launch post, is the gate the company described. The accord sets neither the $2 price nor a ship date.
