Gemini 4 Argon — Google's new flagship model, announced on September 30, 2026. Google is betting on long, multi-step tasks: software development and migration, work with legal and financial documents, cybersecurity, and video analysis. The model can produce responses of up to 1 million tokens, compared with 64,000 in earlier Gemini models. Introductory API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens; after the introductory period ends, it will be $4 and $20 (Google blog). The main caveat: as of October 4, 2026, the model is available only to participants in the Fairwind program for cybersecurity professionals, while paid API customers and Google AI Ultra subscribers will get access later, with no announced date.
This article is for executives, CIOs, and technical leads planning projects with large language models and wanting to know whether it is worth waiting for Argon. We cover capabilities, test results, pricing, access rollout, restrictions for Russian companies, and practical preparation steps. Coverage of other models is provided in separate articles; this text is not a legal assessment of the project.
AIrassvet editorial team. Prepared on October 4, 2026 based on Google’s announcement and Gemini API documentation verified on that date. The editorial team did not test the model — there is no access to it as of publication; the test results are reproduced from Google’s data and identified as such. The article was published on the AIrassvetwebsite, which implements AI solutions for businesses; the services block appears before the conclusion. There are no affiliate links or compensation from Google in this material.
Key takeaways
- Response up to 1 million tokens: Google increased the output limit from 64,000 to 1 million tokens; the input window size was not specified in the announcement.
- Price: introductory — $2 / $10 per 1 million tokens, then $4 / $20; cached input is discounted by 95%. The length of the introductory period was not specified.
- Gradual access rollout: first the Fairwind program and a voluntary pre-release review by the U.S. government, then paid API customers and Google AI Ultra.
- Strengths according to Google’s tests: real software development (77.9% in DeepSWE v1.1), end-to-end business process automation (51.3% in AutomationBench), long video (91.7% in LVBench).
- Russia is not included in the list of regionswhere Gemini API and Google AI Studio are available (Gemini API documentation).
Contents
- What Gemini 4 Argon is
- The main new feature: responses up to 1 million tokens
- Test results
- How much Gemini 4 Argon costs
- How and when access will open
- What Argon means for business use cases
- Availability in Russia and data requirements
- How to prepare for Argon’s launch now
- Frequently asked questions
- Primary sources and verification method
- How AIrassvet helps businesses get ready for new models
- Bottom line: a strong model that cannot yet be bought
What Gemini 4 Argon is
Gemini 4 Argon is Google DeepMind’s frontier language model, built for “long-haul” work: multi-step reasoning, long documents, and tasks the model performs as an agent. The announcement was signed by Koray Kavukcuoglu, Senior Vice President of Google DeepMind and the company’s chief AI architect (Google blog).
Google highlights four areas the model was built for:
- Complex software development — including migration of large codebases.
- Enterprise knowledge work — legal research, financial analysis, taxes.
- Cybersecurity — finding and fixing vulnerabilities.
- Multimodal analysis — understanding text, video, and charts in a single task.
Terms you’ll need below:
- Token — a unit of text used to bill the model; a Russian word usually takes several tokens.
- Output limit — the maximum length of the model’s response per request.
- Context window — the amount of input data the model takes into account in a single request.
- Fairwind — a Google program through which trusted cybersecurity specialists get early access to advanced model capabilities for evaluation (9to5Google).
We covered the background — leadership changes at Google DeepMind and expectations for the Gemini 4 generation — in our article about the Google DeepMind reshuffle.
The main new feature: responses up to 1 million tokens
The output limit for Gemini 4 Argon has increased from 64,000 to 1 million tokens (Google blog). That is about 15.6 times more than earlier Gemini models (calculation: 1,000,000 / 64,000).
Why this matters for business. In the past, a long result had to be assembled from many short responses: the model wrote a chapter, the program stitched the chapters together, and a separate check looked for contradictions between them. With a 1 million token limit, one model can produce in a single request:
- a fully rewritten software module or set of files during migration;
- a detailed analytical report with appendices;
- a full translation or adaptation of a large policy document, without losing links between sections.
Important caveat: Google did not specify the size of Argon’s input context window. Some third-party publications cite 1 million input tokens, but the official announcement does not include that figure. Until the model card appears in the documentation, it is too early to plan tasks around “1 million input tokens.”
The second caveat is cost. A long response is billed at the output token rate: at the introductory price, 1 million output tokens cost $10; after the introductory period, $20. Being able to produce a long response does not mean every response should be long.
Test results
Google published Argon results across several benchmarks. These are vendor-reported figures and have not been independently reproduced:
| Test | What it measures | Argon result |
|---|---|---|
| DeepSWE v1.1 | Real-World Software Development Tasks | 77.9% |
| AutomationBench | End-to-End Business Function Execution | 51.3% (1st place) |
| LVBench | Long-Video Understanding | 91.7% |
| CWE-bench v1 | Code Vulnerability Fixes | 68% (tied for 1st place) |
| Vals Index | Finance, code, law, taxes | index leader |
| Gray Swan IPI | Resistance to indirect prompt injections | test leader |
Source: Google blog.
For comparison with competitors, 9to5Google cites DeepSWE v1.1 results for Claude Opus 5.5 — 74.2% and GPT-6 Astra — 74.1% (9to5Google). Argon's lead is 3.7–3.8 percentage points. That's notable, but not dramatic: on your own tasks, the model ranking may be different.
Two results are especially interesting for enterprise use cases:
- AutomationBench (51.3%) tests end-to-end business function execution. Even the leader scores around half, which means fully autonomous process automation is still not ready to run without human oversight.
- Gray Swan IPI is a test of resistance to indirect prompt injections, where a malicious instruction is hidden in a document or on a web page. For agents that read email and websites, this is one of the main threats; more on that in the article on hacking AI agents through prompt injections.
How much Gemini 4 Argon costs
| Plan | Input, per 1M tokens | Output, per 1M tokens | Cached input |
|---|---|---|---|
| Introductory price | $2 | $10 | 95% discount off the input price |
| After the introductory period | $4 | $20 | 95% discount off the input price |
Source: Google blog. Google did not specify how long the introductory period lasts. As of October 4, 2026, Argon is still not listed on the Gemini API pricing page (pricing page), so details such as batch mode, priority access, and cache storage are still unknown.
Benchmarks for other models on the same date:
| Model | Input / output, per 1M tokens | Source |
|---|---|---|
| Gemini 4 Argon (introductory) | $2 / $10 | |
| Gemini 4 Argon (standard) | $4 / $20 | |
| Gemini 3.8 Flash (paid tier through 12/31/2026) | $0.75 / $3.75 | Gemini API pricing |
| Claude Sonnet 5.5 | $2 / $10 | Anthropic |
| Claude Opus 5.5 | $4 / $20 | Anthropic |
| GPT-6.1 Sol (short context) | $2 / $10 | OpenAI pricing |
Example calculation (Calculated). A contract review task: 200K input tokens and 20K output tokens. At the introductory price: 0.2 × $2 + 0.02 × $10 = $0.60. At the standard price: 0.2 × $4 + 0.02 × $20 = $1.20.
For 1,000 such tasks per month, that's $600 and $1,200 respectively. If you're building your project economics around the introductory price, plan for that price to double.
How and when access will open
Google is rolling out access to Argon in stages and says it's doing so for security reasons (Google blog):
- Fairwind program — trusted cybersecurity professionals; access is opening now.
- Voluntary pre-release review by the U.S. government — running in parallel.
- Paid API customers and Google AI Ultra subscribers — “coming soon,” with no date.
- Developers, businesses, and everyday users — later on.
The first-stage rollout has an important caveat: Fairwind participants are getting the model without cybersecurity guardrails, while Google says the broader release will include the full set of restrictions. For businesses, that means the model’s behavior for early users and everyone else may differ, especially in tasks related to vulnerability analysis.
What Argon means for business use cases
| Task | How Argon can help | What to check before implementation |
|---|---|---|
| Code migration and modernization | Long output, strong DeepSWE results | Quality on your stack, cost of long responses |
| Legal and financial analysis | Leadership in the Vals Index | Accuracy of references to regulations, attorney review |
| Video analysis (quality control, training, safety) | 91.7% in LVBench | Conditions for sending video to the cloud, latency |
| Cybersecurity | CWE-bench results, Fairwind program | Limitations of the public version, usage rules |
| Agents that read external documents | Resistance to indirect prompt injection | Agent permissions, tool isolation |
Separately on video: practical scenarios for analyzing camera footage and meetings are described in the article on agentic video analysis.
Our view: don’t build a plan around a single model. Argon may turn out to be the strongest model for some tasks, but the broad-access date is unknown, the price will double at some point later, and official access is unavailable for Russian companies. The safer approach is to design the system so the model can be swapped out and to verify quality on your own task set.
Availability in Russia and data requirements
Russia is not included in the list of countries and territories where Gemini API and Google AI Studio are available (Gemini API documentation). This applies to all Gemini models, not just Argon.
What this means:
- companies with a foreign legal entity or branch need to check Google’s rules for their jurisdiction;
- transferring personal data to an overseas service is a cross-border transfer under Federal Law No. 152-FZ; see more in the article on 152-FZ and foreign AI;
- for tasks involving confidential data, the practical alternative is models accessible from Russia or models deployed in your own environment; compare options in the overview of the best Chinese and American AI models and in the article cloud or local LLM.
This section is not legal advice.
How to prepare for Argon’s launch now
Waiting for broad access can still be useful. Five steps that will help with any future model:
- Build a benchmark task set. 30–50 real examples from your workflow with correct answers are the foundation of any model comparison.
- Separate the app from the model vendor. A single access layer lets you switch models without rewriting code; see more in the article on a unified LLM API.
- Set up routing. Put complex tasks on the strongest model, and high-volume simple ones on the cheaper model; see the AI model router.
- Run the economics at the standard price. The introductory rate is temporary; the project should still pay off at $4 / $20.
- Prepare a fallback option. If access to the external model disappears, your process should not stop — see the AI service exit plan.
Frequently asked questions
When will Gemini 4 Argon be available to everyone?
Google has not announced a date. Right now, the model is available to Fairwind program participants; next will be paid API customers and Google AI Ultra subscribers, then a broader group of developers, businesses, and users.
How much does Gemini 4 Argon cost in the API?
The introductory price is $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input at a 95% discount. After the introductory period, the price will be $4 and $20. Google did not specify how long the introductory period will last.
What is the context window for Gemini 4 Argon?
Google only specified the output limit — up to 1 million tokens, compared with 64,000 in earlier Gemini models. The input window size was not named in the official announcement; figures from third-party publications should be treated as unconfirmed.
How is Gemini 4 Argon better than Claude Opus 5.5 and GPT-6 Astra?
According to data cited by 9to5Google, in the DeepSWE v1.1 development test, Argon scored 77.9% versus 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. These are vendor results on specific tests; you should verify the ranking on your own tasks.
Can Gemini 4 Argon be used in Russia?
Russia is not among the regions where Gemini API and Google AI Studio operate. Russian companies should consider models accessible from Russia or deploying the model in their own environment.
Should you delay a project until Argon is released?
Usually not. It is better to launch the project on an available model, but design the system so the model can be replaced and keep a benchmark set for comparing new models.
Primary sources and verification method
Facts were verified on October 4, 2026, using the following sources:
- Gemini 4 Argon: Our Next Era of Frontier Intelligence (Google): capabilities, benchmarks, pricing, and access process;
- 9to5Google: comparison with Claude Opus 5.5 and GPT-6 Astra in DeepSWE v1.1, Fairwind overview;
- Gemini API pricing and available regions (Google): pricing and regions;
- Claude Sonnet 5.5 overview (Anthropic) and OpenAI pricing: model prices for comparison.
Methodology: the specifications and benchmark results were taken from the manufacturer’s materials and were not independently reproduced; as of the publication date, the model was not available to the editorial team. Cost calculations were made using public prices. Recommendations are the result of editorial analysis.
How AIrassvet helps you prepare for new models
A new model comes out almost every month, and the companies that win are the ones whose processes and data are ready for quick testing and model replacement. AIrassvet develops AI agents integrated with business systems and deploys language models, including on the customer’s own servers. For the tasks in this article, four areas of work are especially useful:
- process audit and preparation of a benchmark task set for comparing models on quality and cost;
- development of an MVP AI agent integrated with CRM, ERP, or 1C, with a layer that lets you swap models without rewriting the application;
- deployment and fine-tuning of a language model within the company’s environment if foreign APIs are unavailable or do not meet data requirements;
- computer vision and video analysis solutions for production and document workflows.
The first step is to describe one process: current metrics, data sources, constraints, and the criterion you will use to know whether the model works. Discuss the project.
Bottom line: a strong model you still can’t buy
Gemini 4 Argon is Google’s bid to lead in long-context tasks: up to 1 million tokens per response, the company’s best results in software development, business process automation, and video analysis, and an introductory price at the level of mid-tier models. But as of October 4, 2026, the model is available only to a limited group of cybersecurity specialists, no general release date has been announced, and the introductory price will double over time.
For Russian companies, that is compounded by the lack of official access to the Gemini API. So the practical takeaway is simple: do not put projects on hold waiting for Argon. Build them on available models with a swap-ready architecture, and keep a benchmark task set so you can test Argon or any next model within a few days when it becomes possible.