The short answer: Gemini 4 Argon is not yet available as a standard API. Google announced it on September 30, 2026, and made it available only to selected cybersecurity defenders through the Fairwind program, without announcing a date for wider availability. For future budgeting, base your calculations primarily on the price of $4 per million input tokens and $20 per million output tokens, rather than relying solely on the introductory half-price rates.
What Google Actually Announced on September 30, 2026
Google introduced Gemini 4 Argon as a model for long-running tasks in software development, knowledge work, and cyber defense. However, it is not a service that an ordinary business can order immediately and confidently include in a production roadmap.
According to Google’s announcement of September 30, 2026, selected security partners received the first access. Ars Technica describes the same situation: the model has been announced, but a generally available API is still missing.
The most striking figure is the maximum output of one million tokens. The previous limit was 64,000 tokens. A larger output allowance may help with long programming or analytical tasks, but it does not by itself demonstrate the model’s accuracy, reliability, or cost-effectiveness.
For business decisions, three unresolved questions matter most:
- when the API will actually become available,
- what its operational and regional conditions will be,
- whether the model can handle specific business tasks better or more cheaply than current alternatives.
API Pricing: Budget for the $4/$20 Rates
Google announced two pricing tiers. The introductory rates are expected to be $2 per million input tokens and $10 per million output tokens. Once the introductory period ends, prices are expected to rise to $4 and $20 respectively. The announcement does not say how long the discounted period will last.
| Item | Introductory rate | Later rate |
|---|---|---|
| 1 million input tokens | $2 | $4 |
| 1 million output tokens | $10 | $20 |
| 1 million input and 1 million output tokens | $12 | $24 |
The final row is simply the sum of the published rates. It does not include repeated attempts, additional steps in an agentic process, or calls to external tools. Caching has the opposite effect: according to Google, cached input tokens will cost 95% less than standard input tokens.
For example, a task using 250,000 input tokens and 100,000 output tokens would cost $1.50 at the introductory rates:
- input: 0.25 × $2 = $0.50,
- output: 0.10 × $10 = $1.
At the later rates, the same volume would cost $3:
- input: 0.25 × $4 = $1,
- output: 0.10 × $20 = $2.
This is only an example calculation, not an estimate of a typical task. Consumption will depend on the prompt, response length, number of revisions, and the design of the application as a whole.
The introductory price should therefore never be the sole basis for calculating return on investment. The same principle applies when comparing business plans and APIs for other models: what matters is not the cost of one successful call, but the cost of a completed and reviewed case.
What One Million Output Tokens Means
One million tokens is the maximum output limit. It is not a promise that the model will produce a response of that length every time, nor is it evidence that the entire output will be correct and practically useful.
The announced figure also reveals nothing about the size of the input context. Input context and maximum output are two different model properties.
A high limit may be useful for tasks such as:
- processing a large programming task across several consecutive steps,
- producing an extensive specialist analysis,
- operating an agent that must document the process and results of multiple tools.
At the same time, it presents a cost risk. An application may generate unnecessarily long responses, repeat failed steps, or continue working even when it no longer has enough evidence to reach a reliable conclusion.
A business therefore does not need to test whether the model can exhaust its maximum limit. It needs to determine whether the model can complete a specific task at the required quality, length, and cost.
Availability: No Date for a Wider Release
The first limited deployment involves trusted defenders participating in the Fairwind program. Google intends to continue security testing and refining safeguards before a wider release.
According to the announcement, the next stage of access is expected to begin with paying API customers and Google AI Ultra subscribers. However, Google has not provided a specific date.
For an ordinary business, the conclusion is simple: Argon can be included in research and preparations for a pilot, but not in a project that depends on a fixed delivery date.
The available announcements also do not yet confirm practical API limits, regional availability, or the terms governing the processing of business data. Without this information, it is impossible to decide whether the service is suitable for personal data, contracts, customer communications, source code, or other sensitive materials.
How to Prepare a Business Pilot
You can prepare a pilot in advance, but only launch it after gaining access and reviewing the published operating terms.
1. Select Real Tasks
Do not compare models using only generic questions in a chat window. Choose several repeatable use cases, such as:
- fixing a bug in a specific repository,
- reviewing a code change against tests and requirements,
- analyzing a set of technical documents,
- preparing an answer from precisely defined internal sources.
Use synthetic or anonymized data for the first test. Do not upload sensitive business documents to the service until you understand how they will be stored and processed.
2. Define the Correct Outcome in Advance
For each task, specify:
- what the output must contain,
- which errors are unacceptable,
- what a person must verify,
- when the system must stop working,
- when the model must acknowledge that it lacks sufficient evidence.
Without predefined criteria, the test can easily turn into a subjective judgment about which response „feels better."
3. Measure the Entire Process
Recording the tokens used by the final successful call is not enough. You should also track:
- input and output tokens across all steps,
- the number of retries,
- failed or unusable runs,
- the time required for human review,
- human interventions during the task,
- the cost per completed case.
Calculate the cost model using the later rates of $4 per million input tokens and $20 per million output tokens. You can report the introductory price separately, but it should not determine long-term return on investment.
4. Add Safeguards to Your Own Application
The exact design of the future API is not yet known. You should therefore implement basic limits in your own application:
- a spending cap per case,
- a maximum output length,
- a maximum number of retries,
- a list of permitted tools,
- human approval before any irreversible action,
- an audit log of inputs, steps, and results.
If you want to prepare a measurable pilot instead of conducting one-off experiments in a chat, our approach to designing and implementing AI automation in a business covers everything from task selection and security to cost evaluation.
Test Prompt Injection in Your Own Workflow
Google describes Argon as its most resilient model yet against indirect prompt injection. At the same time, it acknowledges that this type of attack requires continuous attention and multiple layers of protection. The vendor’s claim is therefore no substitute for your own security testing.
With indirect prompt injection, the malicious instruction is not contained in the user’s prompt. Instead, it may appear in a document, web page, email, or source code loaded by the agent. The hostile text may attempt to change the task’s objective, request sensitive data, or persuade the agent to use a tool outside the intended workflow.
Your pilot should therefore include a document containing a deliberately malicious instruction and test whether the system:
- separates the user’s instruction from the document’s content,
- refuses an unauthorized action,
- does not send data to another tool,
- records the incident in the audit trail,
- requests human approval.
We describe practical protection layers in our article on AI agent security, permissions, and prompt injection.
Set Data Rules Before Requesting API Access
Before the pilot begins, determine which data may be sent to an external service and which must remain within the company. A general label such as „sensitive" is not enough. At a minimum, distinguish between personal data, business contracts, customer documents, credentials, source code, and internal security information.
Once the API terms are published, you will need to verify in particular how inputs are stored, what is logged, what access the provider has, and whether the data may be used for any additional purposes. If the terms do not meet the requirements of a particular workflow, even a more accurate model will not solve the problem.
You can prepare an initial inventory of tools and data flows by following our guide to AI audits, shadow AI, and data residency.
What You Can Decide Today
Gemini 4 Argon has an intriguing announced output limit, but one million tokens alone is not a business case for adoption. For companies, the deciding factors will be API availability, data-processing terms, performance on their own tasks, and the cost of the entire completed process.
You can already prepare test cases, expected outcomes, security scenarios, and measurement methods. However, you cannot yet base a production deadline or the processing of sensitive data on Argon. A firm date for wider availability and published operating terms are still missing.
Where to find us
ITHOPE is a company based in Brno, Czechia: we do data recovery from disks, RAID and NAS, PC and laptop service and IT for business, all in our own lab at Viniční 3587/216, 615 00 Brno.
Phone +420 775 556 063 (24/7 line). You can bring the device in person, or send it by courier (Zásilkovna, PPL).