Gemini 4 Argon shipped in two safety configurations── Only the contract type and the record of when settings switched over show which configuration a company received; the model name and the published scores do not.

On 30 September 2026, Google announced the AI model Gemini 4 Argon and said that some users would receive it with its cyber guardrails removed. When one AI is handed out under one name with different settings for different recipients, can a company that uses it find out which settings it has? The only clues are the contract type and the record of when the settings switched over; neither the model name nor the published scores will tell you.

01One AI Model Name Now Covers Different Limits for Different Recipients

Whether a company can learn its own settings depends first on the way Google is distributing the model, and Google is distributing it with different settings for different recipients. The first to receive Argon are organisations in a limited programme called Fairwind. The programme began in early September. Governments, Google Cloud customers and security companies take part, and more than 650 organisations had joined when it started.

The version without limits goes to Fairwind participants and to Google's own internal security staff. After Fairwind, Argon goes to companies using it under paid developer contracts and to Google AI Ultra subscribers. No date has been given for general release.

Google has not said which limits it removed. The announcement also does not say how far the unrestricted version will go in its answers.

The single name "Gemini 4 Argon" refers to at least two configurations. A company using the new model cannot tell, from the name or from the scores, whether its own AI is the restricted version or the unrestricted one.

A company that brings the new model into its work therefore needs to put two questions to Google in writing. The first is which type its contract falls under. The second is when its settings switched over. If the company keeps Google's written answers, it can later explain which settings its AI was running with.

02Google Published Two Scores and Left the Lifted Limits Unlisted

One name covers at least two configurations, so which of the two did Google's published scores measure? Google gave scores on two tests.

On a test of finding weaknesses in software, the new model scored 68%. That test is called CWE-bench v1. On DeepSWE v1.1, a test of solving software development tasks, it scored 77.9%.

According to Tech Wire Asia, both figures were measured by Google itself. No outside body produced them.

The announcement does not state which settings were in place when the scores were measured. Reading the coverage, I cannot tell whether the tests ran on the unrestricted version or on the restricted one. Nor does the current reporting include any result from an outside body running the same tests.

There is one published set of scores. The AI that reaches users, however, splits into two kinds according to contract type. Looking at the scores, a company has no means of confirming whether the AI on its own desk is the one that produced them.

03Drug Companies in Japan Already Record Which Material Went to Which Doctor

However closely a company reads the scores, they do not show who received which configuration of the AI. A rule for recording recipients already exists in another industry. Japanese pharmaceutical companies record, doctor by doctor, the materials they used when explaining their products.

In 2018, the Ministry of Health, Labour and Welfare issued its guideline on sales information provision activities for prescription drugs. Under that guideline, a drug company sends its materials for doctors through review by a department independent of the sales division before anyone uses them. What a sales representative told a doctor, down to oral explanations, is written into a work record and kept.

In a drug company, then, which material was used with which doctor, and when, sits in the company's records. It does not depend on one employee's memory.

The drug company case shows that a record of recipients can be built. For the new model, only Google holds the record of which configuration went to whom. A company using the model cannot see Google's record unless it asks for it. Google has not announced that it will show the record to those who ask.

04Google Said a Model This Capable Has to Be Released in Stages

Drug companies keep records of who received their materials. Google, for its part, splits Argon's settings by recipient, and it has given a reason for doing so.

Koray Kavukcuoglu, who leads AI architecture at Google, said that releasing capability at this level safely requires a phased approach. According to an AFP report carried by RTÉ, experts who tried Argon early used it to find weaknesses in hospital software that other frontier models had missed.

The ability to find weaknesses in software helps defenders and attackers alike. Google gave the unrestricted version to defending organisations so that they could use the model's full capability. The Next Web reported that organisations defending against cyberattacks receive the new model first.

The content of the lifted limits has not been published. Because I do not know that content, I cannot judge whether Google's way of distributing the model is sound.

One thing is clear: Google is choosing recipients and changing the settings for them. A company that hands out AI to selected recipients needs a record of who got which version. Without that record, outsiders cannot check afterwards whether the split was carried out as planned.

05A Company Using the New Model Can Check Three Things Itself

Google chooses recipients and changes their settings, and a company on the receiving end can still check three things on its own.

The first thing to check is the contract type. The settings a company receives differ by the contract it uses. There are three kinds of arrangement with Google: membership of the limited programme, a paid developer contract and an Ultra subscription. A company knows which contract it is using. What it cannot see from the announcement is which settings that contract carries, so it confirms that with Google in writing.

The second is the date on which the settings switched over. Because supply widens in stages, the settings can change on a given day even while the contract stays the same. A company that does not know the switch-over date cannot compare last month's answers with this month's under the same conditions.

The third is who measured the scores. The 68% and the 77.9% were Google's own measurements. When a company writes those scores into its internal approval documents, it should add that Google measured them itself.

On the content of its settings, the only thing a user company can keep in its own hands is the written answers it gets from Google. An explanation given orally cannot be verified once the settings have switched over.

06The Material for Choosing an AI Moves from the Name to the Contract Text

Once a company starts checking its contract type, its switch-over date and the source of the scores, the material it uses to choose an AI model changes in three ways.

First, the way companies build comparison tables changes. Tech Wire Asia reported the 68% as Google's own measurement. Placing Google's score next to other vendors' scores gives no guarantee that the tests were run the same way. If a company puts scores into a comparison table, it needs to add a column beside each score showing who measured it.

Second, the assumptions behind internal discussion of AI change. Suppose that in one company, Department A uses the unrestricted version through the limited programme, while Department B uses the restricted version under an ordinary paid contract. The two departments use an AI with the same name, but the settings inside differ. An answer one department gets will not necessarily appear in the other. Google itself has announced that it limits who receives the version without restrictions.

Third, where a company must look for the record changes. In the drug industry, the company that used the material holds the work record. For the new model, only Google knows which configuration went to whom. The only route a company using it has to part of that record is to ask Google in writing. The announcement does not say whether such a request will be answered.

07The Next Dispute May Be Over a Duty to State Per-Contract Settings in Writing

Companies now have to read contract text to learn their settings, and the next issue is whether the law will one day require that disclosure. Two documents offer some guidance.

On 23 September 2026, the attorneys general of 26 US states sent a letter to congressional leadership. They asked for federal oversight of safety testing and standards. They also asked for a system in which, after an incident, investigators can look directly at a company's books and records and then publish the results. And they asked that federal law not override state law.

Article 26 of the EU AI Act requires companies and public bodies that use high-risk AI to keep the logs the AI generates automatically. When police identify a person from recordings after the fact, the Act also requires that each use be written into the police file.

Neither the attorneys general's letter nor the EU AI Act, however, says that settings must be disclosed contract by contract. The letter is a request to Congress, and I have not been able to confirm that it will become law. What the EU AI Act sets is a duty on the side that uses AI to keep records. Article 26 does not address the case in which an AI provider varies its settings by contract.

My own view is this. If AI companies start distributing different settings under one name, an incident investigator will first ask which setting went to whom. If investigators gain the power to look directly at company records, I expect that power to reach the records of how settings were split among recipients.

For a company to know the settings of the AI it is using, it has to look at the contract type and the switch-over record. The model name and the published scores do not reveal the difference in settings.

A company bringing the new model into its work would do well to ask Google three things in writing: its contract type, the switch-over date and the source of the scores. It should keep the written answers. Whether the law will require this disclosure has not yet been decided. Until it is, the written answers a company has kept for itself are the only means it has of explaining the settings of the AI in its hands.