Gemini 4 Argon on Google Cloud: Model ID, $2/$10 Launch Pricing Through Nov. 30, and the 1M Token Limit

Bottom line: Google Cloud lists the model ID gemini-4-argon with launch pricing of $2 input / $10 output per million tokens through November 30, 2026. VentureBeat reports the standard $4 / $20 rate starting December 1. Sources still describe the 1M-token limit in two different ways, so check the console before relying on it.

The callable string and the September 30 press name are no longer the same problem. Google Cloud’s October 13 post, by Michael Gerstenhaber, says enterprise developers can call Argon in the Gemini platform under the model ID gemini-4-argon, and also in Google AI Studio and Google Antigravity. The Google blog the same day prints the introductory price and a November 30, 2026 expiry but does not repeat that model ID. VentureBeat’s October 13 account is the one that says the September 30 announcement never dated the discount, and that December 1 is when the public API price doubles. Three pages from the rollout. One string. Two ways of writing the day the $2 rate ends.

We already wrote the week the public call was not the call. Gemini 4 Argon was unveiled for cyber partners first, and the listed price was not a migration. We already wrote a Sol price that was a different cut from an average saving. Haiku 5.5’s $0.10 line was that cut. We already wrote the earlier card where Opus 5.5 reset prices and Sol halved them. That Sol card is the comparison VentureBeat is now sitting Argon next to. This week’s change is that a Google page printed an ID, and a second Google page printed an end date the September stories did not have.

gemini-4-argon is on the Cloud page. It was not in the September stories.

On October 6 we refused to invent gemini-4-argon as a string you could paste, because the four stories from the unveil did not print a callable ID. Gerstenhaber’s Cloud post now prints it, in backticks, as the model ID in the Gemini platform. That is one vendor page. The October 6 warning was “do not guess.” It was not “ignore the string after Google prints it.”

MarkTechPost’s October 13 writeup covers access paths and a model card without showing the ID. So the string is not “everywhere on October 13.” It is on the Cloud post. If a teammate says the blogs all printed the ID, they have generalized one page. Cite Gerstenhaber for the string. Cite the others for the price and the window.

Android Studio shows up in MarkTechPost’s access list, attributed there to the developer post: Gemini API paid tier in AI Studio and Android Studio.

November 30 is the footnote. December 1 is VentureBeat’s start.

The Google blog’s footnote is the cleanest end date. After the introductory period expires on November 30, 2026, the price of $4 per 1 million input tokens and $20 per 1 million output tokens will apply. The footnote says this applies to the Argon model intended for public access only. That last clause matters. A defender build and a public build were already different doors in the September coverage. The expiry sentence is about the public-access model.

The Cloud post says the introductory rate of $2 per million input tokens and $10 per million output tokens runs through November 30, 2026, with cached input tokens priced at 95 percent off. After November 30, standard pricing of $4 and $20 applies. “Through November 30” and “expires on November 30” are the same boundary if the rate is good for the whole of that day. Neither sentence prints a timezone.

VentureBeat writes the next morning instead of the last night. Unlike the September 30 announcement, which did not specify when the introductory period would end, Google now confirms the reduced pricing expires November 30, 2026. Beginning December 1, the standard public API price doubles to $4 and $20. “Expires November 30” and “beginning December 1” are how you describe one midnight from either side. They are not two price changes. A calendar that ends the discount on November 29 has dropped a day VentureBeat and the footnote both still include. A calendar that starts $4 pricing on November 30 has started a day early relative to “beginning December 1.”

MarkTechPost prints the same pair of rates and the same November 30 line, and then uses the later rate as the comparison: after November 30, $4 and $20 matches Claude Opus 5.5 and sits below GPT-6 Astra’s $10 and $50. That comparison is about the post-intro card. VentureBeat’s comparison, in the same news cycle, is about both cards.

$2 and $10 match Sol in one table, and Opus only after the expiry

VentureBeat’s list prices, which it says exclude tools, infrastructure, and other services:

Argon through November 30 is $2 input and $10 output per million tokens. Claude Sonnet 5.5 is $2 and $10. GPT-6.1 Sol is $2 and $10. Claude Opus 5.5 is $4 and $20. Argon from December 1 is $4 and $20. Meta Muse Wonder is $4 and $20. GPT-6 Astra is $10 and $50. The intro cached-input rate, in VentureBeat’s note under the table, is $0.10 per million tokens.

At the introductory price, VentureBeat says Argon costs half as much per input and output token as Claude Opus 5.5, and one-fifth as much as GPT-6 Astra. Half of $4 is $2. One-fifth of $10 is $2. The output side matches the same way: half of $20 is $10, one-fifth of $50 is $10. From December, the same piece says Argon’s base price moves into the same category as Opus 5.5 and Muse Wonder. “Same category” is their phrase for $4 and $20. It is not a claim that the models cost the same to finish a task. VentureBeat says raw token pricing is only part of the story, because models consume different amounts of reasoning tokens, need different numbers of attempts, and vary in how effectively they complete a task. A router that flips to Argon on December 1 because the sticker matched Opus has used the sticker and skipped that sentence.

MarkTechPost’s table is a different cast. It does not include Sonnet 5.5, Sol, or Muse Wonder. It does include Claude Fable 5.1 at $10 and $50, with a 1 million context and a 128K max output, next to Astra at $10 and $50 with a 1,050,000 context and a 128K max output, and Opus 5.5 at $4 and $20 with a 1 million context and a 128K max output. Argon’s row in that table is $2 and $10 intro, $4 and $20 after November 30, context 1 million combined, max output up to 1 million. The Sonnet and Sol match is VentureBeat’s. The Fable row is MarkTechPost’s. If a slide says “every October 13 table showed Sol at $2,” it has cited a table that did not include Sol.

Cached input is the column both Google pages state as 95 percent off, and only VentureBeat converts into $0.10. Ninety-five percent off $2 is $0.10. That conversion is fair for the introductory input price. The Cloud post states the 95 percent discount in the same sentence as the introductory rate, then says that after November 30 the standard $4 and $20 pricing applies. It does not restate the cache discount on the $4 rate. Five percent of $4 is arithmetic. It is not a line either Google page printed for the post-intro period. If the discount survives, the docs will say so. Until they do, the $0.10 figure is an intro figure.

One million output and one million combined are both in this week’s pages

The September stories, as we wrote them, treated a move from 64,000 output tokens to 1 million output tokens as the product change, and warned that a ceiling is not access. VentureBeat now says that description needs a qualification. The September 30 announcement described a new ceiling of 1 million output tokens, up from 64,000. The new announcement emphasizes a combined input and output limit, supported through a specialized API protocol. Google has not clarified, in the materials VentureBeat is using, how that limit relates to the maximum number of tokens a developer can receive in one uninterrupted response. Witteveen, in that piece, cautions that this specialized capability should not be mistaken for the model’s normal behavior in the Gemini app or other consumer products.

The Google blog’s own phrase is “an industry-leading 1M unified input-output token limit via the Gemini API,” and it points at a continuation protocol in the developer and enterprise docs. MarkTechPost holds both sentences at once. Google raised the output limit to 1 million tokens, up from 64K. The model card, as MarkTechPost describes it, describes a combined input plus output window of up to 1 million tokens, so a long prompt reduces the room left for output. A continuation protocol is documented for long decodes. Those are not three ways of saying “you always get 1 million tokens back.” A combined window means the prompt spends the budget. A continuation protocol means a long decode may not be one uninterrupted response. VentureBeat’s point is that Google has not, in the materials that writer had, spelled out how those two facts meet.

MarkTechPost also says that at introductory pricing, a full 1 million token response costs $10 in output tokens. That arithmetic is $10 per million times one million, which is $10. It is only a bill you can be charged if the API actually emits 1 million output tokens. Under the combined-window reading, a non-empty prompt leaves less than 1 million for the output. Under VentureBeat’s reading, it is not even clear that one uninterrupted response can be that long. Our October 6 note multiplied an extra thousand on that same ceiling. The sticker math, if you fill a million output tokens at $10 per million, is $10.

MarkTechPost’s table also gives Argon a max output of “up to 1M” and gives Astra, Opus 5.5, and Fable 5.1 a max output of 128K. That contrast is the table’s. If Astra’s card does not say 128K, MarkTechPost owns the error, and this piece should not be the place you copy Astra’s cap from. The Argon side of the disagreement is enough: Google blog says unified 1 million, MarkTechPost says output up to 1 million and combined window up to 1 million, VentureBeat says the relationship to one uninterrupted response was not clarified.

The Cloud post does not restate the token window. It does restate a security number that is not a context length. Argon leads the Gray Swan Indirect Prompt Injection benchmark, cutting attack success across 15 adaptive attempts to 0.7 percent. That is Gerstenhaber’s figure for a specific eval, in a paragraph about reading email, documents, and webpages that may contain hidden instructions. It is not a context-window fact, and it is not a reason to skip the prompt-injection controls you already run.

The doors still do not share a noun

VentureBeat’s access paragraph is the widest consumer door in this set. Argon is offered through the paid Gemini API, Google AI Studio, Antigravity, and eligible Gemini Enterprise subscriptions. Google AI Pro and Ultra subscribers can also use the model in the Gemini app and in AI Mode in Search. Antigravity access for individual subscribers requires the more expensive Ultra tier. “Pro gets the app, Ultra gets Antigravity” is VentureBeat’s split. The Cloud post says Argon is available in Gemini for business on web, desktop, and mobile, and that enterprise builders get the API, AI Studio, and Antigravity.

MarkTechPost adds a door the pricing pieces are not about. Cyber defenders get the Fairwind Program, with access to Argon without cyber guardrails. That sentence is MarkTechPost’s account of the developer post. It matches the shape of the September coverage, where defender access and the monitored public path were different builds. A key that can see gemini-4-argon on a paid API is the public door Gerstenhaber described. A key in a defender program without cyber guardrails is the door MarkTechPost still lists separately. Ask which key you have. The model ID being printed does not make those keys the same product.

The Google blog says Argon is rolling out today, globally, and that the rest of the Gemini 4 family is still to come. “Today” in that post is October 13, not a promise that every surface flipped at the same minute. AI Studio, the Gemini app, Antigravity, Android Studio, and Fairwind are five nouns across three writeups. A status page that says “Argon is GA” has picked a noun. The invoice date is easier. Public introductory pricing, on the pages that date it, runs through November 30 and is $4 and $20 after that, with VentureBeat marking the new rate as beginning December 1. The ID to type, on the Cloud post, is gemini-4-argon. The million-token claim still needs a second sentence before you size a prompt around it. Copy the ID. Date the price. Do not size the prompt off a single “1M.”

What this means for you

For budgeting, model two price periods: the launch rate until November 30 and the higher rate from December 1. Set alerts on usage before the switch, and confirm whether the 1M limit applies to output or to input plus output in your own project’s quota page.