Gemini 4 Argon and the New Frontier Balance: Neck-and-Neck Scores, Introductory Prices and Staged Access

On September 30, Google DeepMind announced Gemini 4 Argon, the first model in the Gemini 4 series. It posts strong results in software engineering, enterprise knowledge work and cyber defense. But for enterprises, the most notable part is less the scores than the pricing and the way access is rolled out.

Scores: three models within a point

• 77.9% on DeepSWE v1.1, a long-horizon software engineering test; first place on the Vals Index for enterprise knowledge work at 68.9%.
• On the CWE-bench v1 cybersecurity test, Argon and GPT-6 Astra both scored 68%, Claude Opus 5.5 scored 67%.
• In other words, the top models are now within a point of each other; a single test result is no longer decisive in choosing a model.

Watch the introductory price

Argon's API price is $2 per million input tokens and $10 per million output tokens, matching OpenAI's discounted GPT-6.1 Sol. But this is an introductory price: once the period ends, it rises to $4 and $20, double the launch rate. Cached input is 95% cheaper. A cost estimate built on today's price could be off by half a few months from now.

Access isn't open to everyone

Argon is currently available to a limited group of testers. Google is offering the model without cyber guardrails to more than 650 security organizations enrolled in its Fairwind program, with broader access to follow. So the 'best model' may not be one every company can use today.

Three practical lessons for enterprises

• Plan costs on the standard price, not the introductory one.
• Before switching to a model, check whether access is generally available and what the usage limits are.
• With scores this close, choose based on quality, latency and cost measured on your own workload.

The remot angle

Because model prices and routing rules are managed in one place in remot, redistributing traffic when an introductory period ends or when a new model becomes generally available is a configuration change. Cost and quality measured on real usage decide which model handles which job.