Google’s Gemini 4 Argon now has a 1,048,576-token window, which Google rounds to 1 million tokens, and Malaysian tech coverage including ProductNation Malaysia has picked up the launch. The access rules are the real story. Google states the model is globally available to developers and enterprises through the API, AI Studio and the Gemini platform, but only for paying customers. The Free tier does not include it.
The pricing carries a deadline. Google lists an introductory rate of $2 per 1 million input tokens and $10 per 1 million output tokens through November 30, 2026. From December 1, those rates double to $4 and $20. Developers who build on the model this month will see their bill change within weeks, so the budget math matters as much as the benchmark chatter.
Third-party tools are already moving. Our earlier report on GitLab Duo adding Gemini 4 Argon showed how quickly platform vendors wire a new Google model into their agent stacks. This analysis covers what Google confirmed, what it left out (including any named mention of Malaysia), what the model costs in real workloads, and where it sits against rivals. For more model coverage, see our AI and machine learning section.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Google Announced About Gemini 4 Argon
Google’s announcement names the product Gemini 4 Argon and gives the API identifier gemini-4-argon. The company describes it as globally available for developers and enterprises through three routes: the Gemini API, Google AI Studio and the wider Gemini platform. That is the broadest tier of access, and it requires a paid API account.
Consumers and prosumers get a narrower door. Google says Google AI Ultra subscribers can use the model in Google Antigravity, its agentic development environment. Both Google AI Ultra and Pro subscribers can use it in the Gemini app and in AI Mode in Search. The wording points to a staged rollout in which the most expensive subscription sees the model first in developer tooling.
Google also publishes the technical envelope. The context limit is unified across input and output at 1,048,576 tokens. The standard synchronous API response tops out at 65,536 tokens, while a long-decode mode allows output up to the full 1,048,576. Three thinking levels, low, medium and high, let developers trade latency and cost against reasoning depth. Those four facts shape almost every design decision a team will make with this model.
Does Gemini 4 Argon Reach Malaysia? What Is and Is Not Confirmed
The Malaysian angle needs careful handling. ProductNation Malaysia and other outlets framed the launch for local readers, and the story has circulated as “Gemini 4 Argon reaches Malaysia.” Google’s own statements, however, say “globally available” or list supported regions without naming Malaysia. We could not find a Google statement that singles out the country, so the headline wording is a local framing rather than a confirmed regional launch announcement.
That distinction has practical weight. Global availability through the API usually means a developer with a billing account in a supported region can call the model. It does not guarantee that every consumer surface, such as the Gemini app or AI Mode in Search, appears on the same day in every market, or that a local subscription tier exists at the same price. Readers in Malaysia should check their own Gemini app and AI Studio account rather than assume parity with the United States.
How to check your own access
The fastest test is Google AI Studio. If a paid billing account is attached and gemini-4-argon appears in the model list, the API route is open. For the Gemini app, look for the model in the picker under a Google AI Pro or Ultra subscription. If neither surface shows it, the likeliest causes are a free-tier account or a staged rollout that has not reached the account yet.
The 1 Million Token Context Window in Practice
A 1,048,576-token limit sounds like a clean marketing figure, but the word “unified” changes how teams should read it. Input and output share the same pool. A prompt that fills 1,000,000 tokens leaves only 48,576 tokens for the reply, which is below the 65,536-token synchronous maximum. Teams planning to stuff an entire codebase into a prompt must budget for the answer too.
Rough conversions help. A million tokens covers a very large document set, such as many hundreds of pages of text, or a mid-sized software repository. Exact page and line counts depend on tokenization and language, and Google has not published a conversion for this model, so treat any such figure as an estimate. Code, tables and non-English text typically consume more tokens per visible character than plain English prose.
The window also matters for agents. Long-running coding and research agents accumulate tool output, file contents and intermediate reasoning. A bigger window delays the point where an agent must summarize or discard history, which reduces one of the main failure modes in multi-step work. The cost of that convenience is the input bill, and the cached-input discount exists precisely to soften it.
Gemini 4 Argon Pricing: $2 and $10 Now, $4 and $20 From December 1
Google’s pricing page lists two phases. During the introductory period, which runs through November 30, 2026, input costs $2 per 1 million tokens and output costs $10 per 1 million tokens. Cached input receives a 95% discount, listed at $0.10 per 1 million cached input tokens. Beginning December 1, 2026, the headline rates move to $4 per 1 million input tokens and $20 per 1 million output tokens.
| Item | Introductory (through Nov. 30, 2026) | From Dec. 1, 2026 |
|---|---|---|
| Input, per 1M tokens | $2 | $4 (+100%) |
| Output, per 1M tokens | $10 | $20 (+100%) |
| Cached input, per 1M tokens | $0.10 (95% discount) | Not stated in Google’s listing |
| Free tier access | No | No |
| Maximum synchronous output | 65,536 tokens | 65,536 tokens |
The headline reads as a 100% increase on both input and output, and the arithmetic confirms it. Anyone who signs a budget in October at the introductory rate without modeling December will see a doubling. Google has not said whether the cached-input rate also changes after the introductory period, so teams that rely heavily on caching should treat the $0.10 figure as temporary until the pricing page says otherwise.
For context on how Google’s caching economics compare with OpenAI and Anthropic, our breakdown of Gemini, ChatGPT and Claude API cache pricing shows why a 95% cache discount can matter more than the headline per-token rate for repeated-context workloads.
What a Real Request Costs Before and After the Price Change
List prices only become useful once they meet a workload. The table below applies Google’s published rates to five request shapes. These are our own calculations from the listed prices, not Google figures, and they ignore any taxes, regional pricing or discounts that may apply to a given account.
| Request shape | Intro price | From Dec. 1 |
|---|---|---|
| 50,000 tokens in, 2,000 out | $0.12 | $0.24 |
| 200,000 tokens in, 8,000 out | $0.48 | $0.96 |
| 900,000 tokens in, 65,536 out | $2.46 | $4.91 |
| 1,000,000 tokens in, 40,000 out | $2.40 | $4.80 |
| 900,000 cached tokens in, 65,536 out | $0.75 | Not stated |
Two patterns stand out. First, a single maxed-out request costs a few dollars, which is cheap for a one-off analysis but adds up fast for an agent that makes hundreds of such calls per day. Second, the cached row shows the lever that matters. Reusing a 900,000-token context at the cached rate cuts that request from $2.46 to $0.75 during the introductory period, a drop of about 70% because output tokens still bill at the full rate.
A small estimator you can run
INTRO = {"input": 2.00, "output": 10.00}
FROM_DEC_1 = {"input": 4.00, "output": 20.00}
def cost(rates, input_tokens, output_tokens):
total = input_tokens * rates["input"] + output_tokens * rates["output"]
return total / 1_000_000
for name, rates in [("intro", INTRO), ("from Dec 1", FROM_DEC_1)]:
print(name, round(cost(rates, 900_000, 65_536), 2))
# intro 2.46
# from Dec 1 4.91Swap in your own average prompt and reply sizes, then multiply by daily call volume. Teams that skip this step usually discover the doubling on the first December invoice.
Who Can Use Gemini 4 Argon Today
Access splits along subscription lines. The table maps each surface to the audience Google names, plus the one item Google does not name.
| Surface | Who gets access | Status |
|---|---|---|
| Gemini API, AI Studio, Gemini platform | Developers and enterprises, paid API customers | Globally available per Google |
| Google Antigravity | Google AI Ultra subscribers | Available |
| Gemini app | Google AI Ultra and Pro subscribers | Available |
| AI Mode in Search | Google AI Ultra and Pro subscribers | Available |
| Free API tier | Nobody | Not available |
| Malaysia (named) | Not specified by Google | Unconfirmed |
The structure rewards the top of the stack. Ultra subscribers reach the model in three places, Pro subscribers in two, and free users in none. Google used a similar ladder before, and our report on Gemini 3.6 Flash arriving in Antigravity described how the company routes new models through its developer environment first.
Why the Free Tier Is Out
Google’s documentation states that Gemini 4 Argon is not available on the Free tier. The company has not published its reasoning, so any explanation is inference. The likeliest driver is compute. A model that accepts 1 million tokens per request is expensive to serve, and a free tier invites bulk experimentation, scraping-style usage and abuse that a paid billing relationship discourages.
The business logic runs in a second direction too. A paid wall turns developer curiosity into a billing relationship before the introductory pricing ends. By December 1, any team that has already integrated the model and shipped a feature on it faces switching costs, which makes the doubled rate easier to absorb. That is analysis, not a Google statement, but it fits the sequence of a low introductory price followed by a step up.
Hobbyists and students lose the most. Earlier Gemini generations let anyone prototype against the API at no cost. With Argon, the entry ticket is a billing account, even if the introductory rates keep small experiments cheap. A test that uses 50,000 tokens in and 2,000 out costs about 12 cents at the introductory rate, so the barrier is more about friction than money.
Thinking Levels: Low, Medium and High
Google exposes three thinking levels: low, medium and high. The fact sheet does not list a separate price per level, and Google has not published how many reasoning tokens each setting typically consumes, so teams should measure rather than guess. Reasoning tokens generally count toward output, which means a higher level can raise the bill on the same prompt even when the visible answer stays short.
A sensible default is to start at low for classification, extraction and routing tasks, then move to medium for code review and summarization of long documents. Reserve high for multi-step planning, hard debugging and legal or financial analysis where an error costs more than extra tokens. Logging reasoning-token counts per request for the first week gives a defensible cost curve before the December rate change.
Output Limits: 65,536 Tokens by Default, 1,048,576 With Long Decode
The standard synchronous API maximum is 65,536 tokens. Google separately describes output of up to 1,048,576 tokens with long decode. The two figures serve different jobs. The synchronous cap suits chat, code generation and typical document tasks. Long decode targets rare cases such as generating a full book-length draft, a large dataset or an entire code module in one pass.
Because the window is unified, a 1,048,576-token output leaves no room for input. In practice, any long-decode job pairs a short prompt with a very long reply. At the introductory output rate of $10 per 1 million tokens, a near-maximum reply costs roughly $10.49 before input charges. After December 1, the same reply costs about $20.97. Teams should reserve this mode for work that truly needs one continuous output.
Where Google Stands in the Frontier Model Race
Gemini 4 Argon lands in a crowded field. Our LLM leaderboard coverage compared Claude Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon and found Opus 5.5 ahead of Gemini by 1.7 points on that board. A gap that small sits inside the noise of most evaluations, which is why pricing, context and tooling now drive adoption more than a single leaderboard rank.
Google’s edge is distribution. The same model reaches developers through the API, enterprises through the Gemini platform, power users through Antigravity and the broader public through the Gemini app and AI Mode in Search. No rival owns a search product with that reach. Our look at chatbot usage share across Claude, ChatGPT, Gemini and Grok shows why that distribution advantage matters when the model quality is close.
Competitive Comparison: Argon Against Rivals With Big Windows
Million-token context is no longer exclusive to Google. We covered Kimi K3.1 Pro, which Moonshot AI launched with a 1M context and open weights. That puts a different kind of pressure on Google. A team with GPU capacity can self-host an open-weight long-context model and pay for hardware instead of per-token rates, while Argon offers managed reliability and the Google ecosystem.
The comparison therefore turns on three variables rather than one. Managed models trade higher per-token cost for convenience, uptime and integration. Open-weight models trade engineering effort and infrastructure for control and a flatter cost curve. Closed rivals from Anthropic and OpenAI compete on benchmark quality and agent behavior. We have not verified per-token prices for those rivals in this run, so we do not quote them here. Readers should check each vendor’s current pricing page before deciding.
Historical Context: From Small Windows to a Million Tokens
Long context has been a Google selling point across several Gemini generations, and the company has repeatedly used window size to separate its models from competitors. Early large language models handled a few thousand tokens. The jump to hundreds of thousands, and then to a million, changed what developers attempt. Whole repositories, long contracts and hours of transcribed audio now fit in one request.
The pricing pattern is newer. Many AI vendors have courted developers with generous free tiers. Argon’s structure runs the other way, with no free tier and a scheduled step up after the introductory period. That suggests the industry is shifting from land-grab pricing toward margin discipline, driven by the cost of serving long-context, reasoning-heavy models at scale.
Market Impact for Developers and Enterprises
For developers, the immediate effect is a planning deadline. Anything shipped on the introductory rate should carry a December 1 cost review. Product teams that sell AI features at a fixed subscription price face margin compression if their usage is input-heavy, since input tokens double from $2 to $4 per million. Caching and prompt trimming become first-order engineering tasks.
For enterprises, the global availability claim plus the Gemini platform route makes Argon easy to pilot inside existing Google Cloud relationships. Procurement teams should still ask three questions. Does the contract lock the introductory rate? Which regions can serve the model? What data-handling terms apply to long documents? Google’s public statements do not answer these, so buyers need written answers from their account teams.
In Malaysia and the wider region, the effect depends on the unconfirmed regional details. Developers there can likely reach the API through a supported billing account, yet consumer features may roll out on a different timeline. Local startups building document-processing and customer-service tools in Malay, English and Chinese should test token consumption in those languages, since tokenization varies and costs can differ from English benchmarks.
Five Predictions for Gemini 4 Argon
These are our analysis, not reported facts, and each is a judgment call that could miss.
- The December 1 step-up will stick. Google has already published the $4 and $20 rates, and reversing a scheduled increase before it takes effect would be unusual for a model with strong early demand.
- Caching will become the headline feature. With a 95% discount at the introductory rate, teams that reuse large contexts will build around it, and Google will likely extend or reprice it after November.
- A cheaper sibling will follow. Google has historically shipped lighter, cheaper variants after a flagship, and the no-free-tier policy leaves room for a lower-cost model to cover hobbyists.
- Regional rollout details will clarify within weeks. The Malaysia question exists because Google’s statements are broad. Local reporting and developer posts will likely settle which consumer features are live in each market.
- Rivals will answer on price, not window size. A 1 million token window no longer differentiates by itself, so competitors will target the post-December rate with promotions or larger cache discounts.
What Teams Should Do Before December 1
Start with measurement. Log input tokens, output tokens, reasoning level and cache hits for every Argon call during the introductory period. That data converts a vague worry about doubling into a number per feature. Then model the December cost using the $4 and $20 rates and compare it with alternatives for each workload.
Next, cut waste. Trim system prompts, deduplicate retrieved passages and move stable context into the cache. Route simple tasks to low thinking, and keep high for the minority of requests that need it. Finally, keep a fallback model wired in, so a pricing change or regional gap does not leave a product without an AI backend.
FAQ: Gemini 4 Argon Access, Pricing and Limits
Is Gemini 4 Argon available in Malaysia?
Google says the model is globally available through the API, AI Studio and the Gemini platform, but its statements do not name Malaysia. ProductNation Malaysia and other outlets framed the launch locally. Check AI Studio or the Gemini app on your own account to confirm.
How much does Gemini 4 Argon cost?
Through November 30, 2026, Google lists $2 per 1 million input tokens and $10 per 1 million output tokens. Cached input costs $0.10 per 1 million tokens. From December 1, 2026, the rates become $4 and $20.
Can I use Gemini 4 Argon for free?
No. Google states the model is not available on the Free tier. Access requires a paid API account or a Google AI Ultra or Pro subscription, depending on the surface.
What is the Gemini 4 Argon context window?
The unified input and output limit is 1,048,576 tokens, which Google describes as 1 million tokens. Because input and output share that pool, a very long prompt leaves less room for the reply.
How long can a single response be?
The standard synchronous API maximum is 65,536 tokens. With long decode, Google says output can reach 1,048,576 tokens.
Which subscriptions include Gemini 4 Argon?
Google AI Ultra subscribers get it in Google Antigravity. Google AI Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search. Paid API customers get it through the API.
What is the API model name?
The identifier is gemini-4-argon. Thinking levels are low, medium and high.
Will the cached-input discount last after November 30?
Google lists the 95% discount, $0.10 per 1 million cached input tokens, during the introductory period. It has not stated the post-December cached rate in the material we reviewed, so plan for uncertainty.
