Gemini 4 Argon launches with limited access and results that are still hard to compare

A model built for long, complex tasks

Google DeepMind introduced Gemini 4 Argon as an AI model aimed at extended coding, professional and cybersecurity tasks. The company highlights its ability to sustain multi-step workflows and generate long responses. However, the available information does not independently verify every capability Google attributes to it.

Among the figures released is an output limit of up to one million tokens, compared with 64,000 tokens attributed to earlier Gemini models. That limit describes how much text the model can produce in a response; it does not necessarily indicate how much material it can read, nor does it guarantee that it will complete a task accurately. Google has also reported results on coding and security tests, but those scores should be understood as outcomes on specific evaluations, not as a universal ranking.

Initial access is limited to selected specialists

Argon’s initial rollout is restricted to selected cybersecurity specialists and organizations through Google’s Fairwind program. The company links this cautious approach to the model’s claimed ability to analyze and fix vulnerabilities, and to the need to evaluate its safeguards before widening access.

Google has not set a specific date for general availability. The information provided describes a possible later expansion to paid API customers and Google AI Ultra subscribers, but gives no confirmed schedule. The expected sequence should therefore not be read as a launch date or a guarantee that every feature will be available to each group.

Benchmarks offer signals, not a definitive ranking

Google attributes a score of 77.9% to Argon on DeepSWE v1.1, an evaluation of long-horizon software engineering. In the same comparison, the cited figures for GPT-6 Astra and Claude Opus 5.5 are lower. Results have also been reported on cybersecurity and enterprise-work evaluations. These figures describe specific tests; they do not show that Argon is better across every kind of coding, analysis or everyday use.

The external comparisons in the available material are not fully consistent. One account attributes a score of 53 for both Argon and GPT-6 Astra to Artificial Analysis’s Intelligence Index; another evaluation in the supplied documentation does not make it possible to verify the original result or its methodology. The complete protocols, exact versions, prompts and tool budgets needed to reproduce all the comparisons have not been provided either. Argon’s relative position against its competitors therefore remains unsettled.

Token consumption changes the cost per task

The prices reported for Argon are $2 per million input tokens and $10 per million output tokens during an introductory period. Prices have also been said to rise to $4 and $20, respectively, after that period, with a discount for cached input. The available information does not confirm when the introductory pricing will end or independently verify the full commercial terms.

The price per token is not enough to compare effective costs. A model that needs more tokens, multiple attempts or additional supervision may cost more to complete the same task, even if its unit price is lower. The available sources include a secondary claim that Argon used more than twice as many tokens per task as GPT-6 Astra in certain tests; the original records needed to validate that claim have not been provided. It should be treated as a point to investigate, not as a general property of the model.

Cybersecurity helps explain the cautious approach

Google presents Argon as capable of identifying, validating and helping fix software vulnerabilities. That capability could be useful to defensive teams, but it also calls for an assessment of misuse risks. Restricted distribution and testing with selected participants have been described as preventive measures; by themselves, they do not demonstrate that the model is safe in every environment or against every attack.

The available information mentions planned safeguards against malicious requests, prompt-injection attacks and unwanted behavior, as well as isolated environments for certain evaluations. No detailed technical results have been provided to measure how effective each control is. Claims about security and resistance should therefore be understood as statements about the design or tests described, not as guarantees of protection.

What to verify before choosing it

Google’s figures, results attributed to external evaluators and comments about user experience do not carry the same evidentiary weight. A sound comparison should specify the model version, date, tasks, prompts, tool access, token limits and scoring method. Without those details, apparently small differences may reflect different conditions rather than a real change in capability.

For organizations considering Argon, the most useful test is an internal evaluation using representative tasks and criteria set before the comparison. It should cover accuracy, success rate, human corrections, latency, total consumption and security controls. Until suitable access and comparable results are available, the prudent description is that Argon is a Google proposal with limited access and promising early signals on some tests—not a model that has generally outperformed GPT-6 Astra or Claude Opus 5.5.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.