Gemini 4 Argon: Pricing, Access and 1M Tokens

Google has moved Gemini 4 Argon from a limited early-access model to broad commercial availability. Since October 13, 2026, developers and enterprises can use it through the paid Gemini API, while Google AI Pro and Ultra subscribers can select it in the Gemini app and AI Mode in Search. Ultra subscribers also receive access through Google Antigravity.

The launch is aimed at difficult work rather than quick everyday questions. Google positions Argon for long-running software engineering, multimodal document analysis, science, finance, legal research and defensive cybersecurity. Its headline specification is a unified input-and-output limit of 1,048,576 tokens, but access, pricing and feature support differ sharply between the consumer app and the API.

Gemini 4 Argon at a glance

Availability: globally available from October 13, 2026, through the paid Gemini API and selected Google AI subscription products.
Consumer access: Google AI Pro and Ultra subscribers can use Argon in the Gemini app and AI Mode in Search.
Developer access: the stable API model ID is gemini-4-argon in Google AI Studio and the Gemini API paid tier.
Context and output: a unified 1,048,576-token limit, with synchronous output capped at 65,536 tokens unless long decode is used.
Introductory API price: $2 per million input tokens and $10 per million output tokens through November 30, 2026.
Standard price beginning December 1: $4 per million input tokens and $20 per million output tokens.
Supported inputs: text, images, video, audio and PDF; output is text.
Important omissions: no image generation, audio generation or Live API support on this model.

Who can use Gemini 4 Argon now

The simplest route is the Gemini app. Google AI Pro and Ultra subscribers can choose Argon for complex prompts without configuring an API. The same subscriber tiers can use it in AI Mode in Search, where web grounding and follow-up exploration are part of the interface. Free Gemini accounts are not listed for Argon access in Google’s launch announcement.

Developers can call the stable gemini-4-argon model through the paid Gemini API in Google AI Studio. Google also lists enterprise access through its business products and Gemini platform. Antigravity, Google’s agentic coding environment, is narrower: the announcement identifies Google AI Ultra subscribers for consumer access there, while enterprises have a separate path.

Availability does not mean unlimited use. Consumer plans can apply usage caps, and API customers pay for tokens and any other billable services. Teams should confirm the plan, region and account policy shown in their own Google console before committing a production workflow.

What Gemini 4 Argon is built to do

Argon is Google’s flagship reasoning model for tasks that require many connected steps and substantial context. Google highlights large-codebase refactoring, professional document analysis, scientific reasoning, autonomous agents and defensive security work. The model can accept text, images, video, audio and PDFs, which makes it useful when a project mixes source code, specifications, diagrams, recordings and long reference documents.

Its tool support includes code execution, function calling, file search, URL context, structured outputs, Search grounding and Google Maps grounding. Computer use is available in preview. Thinking can be set to low, medium or high, giving developers a way to trade speed and cost against deeper reasoning.

Argon is not a universal media model. It returns text and does not support image generation, audio generation or the Live API. Teams that need real-time voice interaction or generated media will still need another Gemini model or a multi-model workflow.

The one-million-token limit needs careful reading

Google describes Argon as having an industry-leading one-million-token unified input-output limit. Unified means the input and generated response share the same overall budget. A very large prompt therefore leaves less space for the answer, and a long answer leaves less room for source material.

The API reference lists a default synchronous maximum output of 65,536 tokens. Reaching output lengths closer to the full one-million-token ceiling requires Google’s long-decode continuation workflow. Developers should not assume that a normal single request will return a book-length response merely because the headline limit is one million.

A large context window also does not guarantee perfect recall. Important instructions can be diluted by redundant files, conflicting policies or repeated drafts. Strong workflows still need a clear task, organized source material, explicit output requirements and validation against the original evidence.

Gemini 4 Argon pricing before and after November 30

The launch price is temporary. Through November 30, 2026, Google charges $2 per million input tokens and $10 per million output tokens for the public Argon model. Cached input receives a 95 percent discount, although the model page says Argon supports implicit caching rather than explicit caching.

Beginning December 1, the published rate doubles to $4 per million input tokens and $20 per million output tokens. That date matters for pilots: a workflow that looks affordable in October or November can cost twice as much at the same token volume in December.

For a simple illustration, one million input tokens plus 100,000 output tokens would cost about $3 during the introductory period: $2 for input and $1 for output. At the standard rate, the same uncached request would cost about $6. This example excludes other platform charges and assumes the published public-model rates.

Measure real prompt and output sizes before estimating a monthly budget.
Remove duplicate files and stale context instead of paying to resend them.
Test whether a lower thinking level is sufficient for routine work.
Track the December 1 price change separately from usage growth.
Use production limits, logging and human approval for costly agent loops.

Gemini app, AI Mode, API and Antigravity are different products

In the Gemini app, Argon is packaged as a model choice for subscribers. Users provide prompts and files through Google’s interface, while plan limits and product settings control availability. AI Mode in Search adds a research-oriented surface that can ground responses in current web results.

The API gives developers much more control over tools, thinking level, structured outputs and application logic, but it also creates responsibility for authentication, budgets, logging, safety and data handling. Antigravity is designed for agentic coding and longer software tasks. It is not simply another name for the Gemini chat interface.

A team should choose the surface around the job. Occasional analysis may fit the Gemini app. A repeatable business workflow belongs in the API only after the team has tested costs, permissions and failure handling. Large autonomous coding tasks may justify Antigravity, but they still require review before code is merged or deployed.

Capabilities developers should verify before migrating

Multimodal inputs are supported, but generated output is text only.
Code execution, file search, function calling and structured outputs are supported.
Search and Google Maps grounding are supported; URL context is also available.
Computer use is marked preview and should not be treated as a fully stable production dependency.
Implicit caching is supported, while explicit caching is not listed for Argon.
Batch, Flex and Priority inference are listed as consumption options.
Live API, image generation and audio generation are not supported.

These differences can break an application that swaps only the model ID. A system designed for explicit caches, live audio or generated images needs architectural changes rather than a one-line migration. Google’s latest-model guide is the correct place to check request changes and continuation behavior before production use.

How to interpret Google’s performance claims

Google reports that Argon ranks first on the Vals Index and cites internal or partner results across coding, science and cybersecurity. Examples include faster codebase refactoring, improved framework performance and assistance with vulnerability discovery. These are useful signals, but they are not a promise that every customer will reproduce the same gains.

Workload design matters. A model can perform strongly on a curated benchmark while struggling with an organization’s private terminology, incomplete documentation or unusual approval process. The safest evaluation uses representative tasks, blinded scoring where practical and a fixed comparison against the current system.

Google also reports improvements from internal Antigravity use, including more agentic code submissions and fewer rollbacks. Because those figures come from Google’s own environment, buyers should treat them as launch evidence rather than an independent service-level guarantee. Reuters separately reported the wider October 13 commercial launch and its focus on developers and businesses.

Safety and cybersecurity limits still matter

Argon’s ability to inspect code and work across tools raises the consequences of a bad instruction. Google says the model received new jailbreak defenses, prompt-injection training and agent monitoring. The company reports a 0.7 percent attack success rate on a multi-attempt indirect prompt-injection benchmark and says shortcut attempts fell to 2 percent in selected coding evaluations.

Those are vendor-reported evaluation results, not proof that prompt injection is solved. A production agent can still encounter malicious instructions in webpages, email, documents or source repositories. Tool permissions, network access and approval gates should be limited to the smallest scope the workflow needs.

For consequential changes, keep a human checkpoint before deployment, deletion, permission changes, purchases or messages to third parties. Security teams should also separate defensive testing from unrestricted exploit generation and follow the access conditions of Google’s Fairwind Program.

Should existing Gemini users switch

Subscribers should choose Argon when the task is genuinely difficult: a large repository, a long multimodal evidence set, a multi-stage research problem or an analysis that benefits from extended reasoning. It is likely excessive for rewriting a short email, summarizing one page or answering a simple factual question.

API users should run a controlled trial before migrating traffic. Compare quality, latency, token use, tool reliability and total cost with the current model. Recalculate the budget at the December rates, not only the introductory price. If a smaller model meets the quality target, routing easy requests away from Argon can reduce cost and delay.

Organizations in legal, financial, medical or security-sensitive work should keep domain review. Google describes strong professional capabilities, but the output remains generated text and can contain mistakes, unsupported conclusions or outdated assumptions.

A practical first-week test plan

Choose five to ten real tasks that are difficult for the current model.
Define success criteria before testing, including factual accuracy, completeness, latency and cost.
Run the same inputs through Argon and the current baseline without changing the rubric.
Inspect citations, calculations, code changes and tool actions against primary evidence.
Record token use under both introductory and December pricing.
Keep permissions read-only during early agent tests whenever possible.
Move only the workloads with a meaningful, repeatable improvement.

A relevant resource for AI-assisted coding

Readers building their first agentic development workflow can look up AI Vibe Coding by Jordan R. Ashford on Amazon. The book’s verified ISBN is 9798253607541; confirm the seller, edition and price before ordering.

As an Amazon Associate, ReaderSpice earns from qualifying purchases.

Bottom line

Gemini 4 Argon is now a real option for paid subscribers, developers and enterprises, with a combination of deep reasoning, multimodal input, broad tool support and an unusually large unified token budget. The most important limits are easy to miss: normal synchronous output is much smaller than the headline ceiling, several media and real-time features are unsupported, and the introductory API price ends after November 30.

Use Argon for work that earns the extra reasoning and context. Test it on representative tasks, validate the output, keep agent permissions narrow and budget at the December rate before making it a default. For simple prompts, a faster and less expensive model may remain the better choice.