Google introduced Gemini 4 Argon on September 30, 2026, a new model for extended professional tasks, software development and cyber defense. Initial access is being granted to a selected group of partners in the Fairwind program. Despite the broad announcement, this is not yet an opening to every Gemini user, and an existing subscription does not guarantee access to the model.
The company reports an output limit of a million tokens, compared with 64 thousand previously. These are tokens the model can generate during its work, not simply the amount of material that can be fed into it. The difference matters: the ability to read a large document and the ability to keep generating output throughout a task are separate capabilities, with different implications for working time and the bill.
The results table shows advantages, and gaps too
Google DeepMind's model page presents results across a range of tests. In DeepSWE v1.1, which covers software development tasks, Argon scored 77.9%, compared with 74.1% for GPT-6 Astra and 74.2% for Claude Opus 5.5. This is a comparison published by Google, not an independent AI24 test or a promise of the same gap in every code repository.
The same table presents a different picture in FrontierSWE v2: Argon scored 55.0%, compared with 65.5% for GPT-6 Astra and 62.3% for Claude Opus 5.5. In Terminal-bench 4.0, too, Argon's result of 57.4% is below Claude Opus 5.5's 66.4%. Calling a model a flagship does not turn every row in the table into a win.
CWE-bench v1 lists a result of 68.0% for both Argon and GPT-6 Astra, while Google reports 51.3% for Argon in AutomationBench. These tests examine different kinds of work, so the percentages should not be combined into an improvised average or translated directly into the share of tasks a particular office could hand over to AI.
A useful comparison begins by asking exactly what was tested. Fixing a bug in a familiar repository, using tools in a defined environment and analyzing a business document are not identical tasks. A team evaluating a model needs to identify which test resembles its work, then check whether the operating conditions suit its system.
It is also worth separating a single impressive result from reliability across a sequence of actions. If a task requires reading material, editing files and running tests, success at the first stage does not complete it. The business measure is the usable deliverable and the time needed to reach it, including human review time.
A small comparative experiment can include a task the existing system handles well, one it struggles with and a case in which essential information is missing. This tests not only the ability to make progress, but also the ability to recognize when there is no basis for an answer. The examples should be kept fixed so that a difference in the result does not come from changing the question.
It is important to compare the permissions granted in each run as well. A model given access to additional information or different tools is not being tested under the same conditions. This distinction does not negate the value of public benchmarks; it defines what needs to be known before transferring their conclusions to a local system.
Why access begins with cyber defenders
Google DeepMind's Fairwind program is intended to give selected organizations early access to defensive capabilities. Its page lists governments, healthcare providers and telecommunications infrastructure among its target audiences. The stated rationale is to enable defenders to find and fix vulnerabilities before equivalent capabilities are used against them.
According to the program, some partners receive access to Argon and can use it within CodeMender, a dedicated agent for finding vulnerabilities and fixing code. Participation includes suitability checks and usage rules. Dual-use tasks, such as malware analysis and authorized attack simulation, are permitted within defined defense and research purposes.
There is a distinction here between the model's power and the user's right to operate it. The ability to analyze a system does not grant permission to test another party's system. Even within a formal program, it is necessary to define which assets are in scope, who authorizes the actions and how sensitive findings are handled.
For an Israeli organization involved in software defense, this could be a relevant avenue for an eligibility inquiry. The publication does not confirm that every Israeli company will be accepted, guarantee an admission date or promise local service. Lack of access in an ordinary consumer account does not by itself indicate a fault.
In practical terms, it is also better to distinguish between a proposed fix and one that has been deployed. A security test may produce code that looks convincing but changes existing behavior. An evaluation process should also include reproducing the problem, testing the change and demonstrating that the system's normal operations continue to work.
A million tokens are not an instruction to produce a million tokens
Long output can help when a complex sequence of work needs to be maintained. It can also make oversight harder if the user receives a large deliverable that nobody knows how to check. The key question, therefore, is not how much the model can write, but at which points it is possible to verify that the work is progressing correctly.
One possible example is migrating a software component to another language. The work can be divided into mapping behavior, making a small change and checking compatibility. That allows the result to be compared at every stage. A single extended run may be convenient, but it does not replace defining what must not change.
When working with documents, too, it is advisable to request a focused result: which data was requested, which source it came from and what remains unknown. A longer answer is not necessarily a more complete one. Sometimes a short summary that clearly distinguishes facts from gaps makes a better decision possible.
Introductory pricing and availability require separate checks
Google announced introductory pricing of $2 per million input tokens and $10 per million output tokens. After the introductory period, the stated prices are $4 and $20 respectively. The announcement does not specify an end date for that period, so a long-term budget should not rely solely on the discounted rate.
These are model usage rates, not a monthly subscription price or a commitment to the cost of an entire task. Comparing costs requires measuring how much input and output the same work consumes, how often it is repeated and how many corrections it needs. Comparing prices for a million tokens alone leaves part of the bill out of the picture.
The announcement did not publish dedicated Hebrew results. When suitable access opens, a local test should include Hebrew documents, names and terms the team actually uses, and outputs that can be verified. At present, the announcement provides a new model and a limited distribution program. A decision to move real work to it will require both actual access and evidence from the organization's own tasks.