Gemini 4 Argon: Google's new model at half price for a limited time

What's new

Google has launched a new AI model, Gemini 4 Argon. It was announced on September 30, 2026, but initially rolled out only to a small group: trusted cyber defenders in the Fairwind program. According to Google's announcement, broader access for developers, businesses and consumers is coming "soon," and paying API customers and Google AI Ultra subscribers are set to get it first.

According to the well-known AI blogger Simon Willison, the model was available on October 13 through the Gemini API (the access developers use to build the model into their own applications) and for customers on Google's $100-a-month plan. He has also updated his tool llm-gemini to version 0.35 so that it works with Argon. Whether the model can be used in Denmark or the EU is not stated in the sources.

Cheaper or better?

Price is the big talking point. During the introductory period, Argon costs $2 per million input tokens and $10 per million output tokens. Tokens are the small chunks of text that AI models read and write in. If the model stores input in advance (so-called cached tokens), the price is 95 percent lower during the period.

After the introductory period, the standard price is $4 per million input tokens and $20 per million output tokens. Willison writes that the price doubles on December 1. Google's own announcement mentions no date, only an "introductory period."

For comparison, Willison says the launch price is the same as GPT-6 Sol, half that of Claude Opus 5.5 and one-fifth that of Claude Fable 5.1 and GPT-6 Astra. After the doubling, Argon will be on par with Claude Opus 5.5 but still slightly cheaper than GPT-6 Astra, he estimates. A monthly price for regular users beyond the $100 plan has not been disclosed.

Willison does point out, however, that token prices can no longer be compared directly. What matters is how many tokens a model uses to reach a goal when it works independently as an agent. According to Artificial Analysis, which Willison cites, it costs $2,407 to run their test suite with Argon. That is close to GLM-5.3 and significantly cheaper than the latest models from OpenAI, Anthropic and Meta.

What's smart

Argon can deliver up to 1 million output tokens in a single response. Google's text mentions a previous limit of 64,000, so this is a significant boost for long tasks, such as large coding projects or long reports.

Google's own benchmarks show: 77.9 percent on DeepSWE v1.1 (software tasks), 51.3 percent on AutomationBench (first place), 68 percent on CWE-bench v1 (shared first place, vulnerabilities in code) and 91.7 percent on LVBench. These are the vendor's own figures, so they should be taken with a grain of salt.

Willison, who usually tests models by asking them to draw a pelican as an SVG image, rates Argon's drawings as excellent at low, medium and high settings. A high-setting pelican cost him 17.166 cents.

What it's good for

Google highlights three areas: software development, knowledge work in businesses such as law and finance, and cyber defense. That fits with the model first being given to security professionals.

In practice, this means developers and businesses with heavy, self-running tasks can get a powerful model at a low price, as long as the introductory price applies. Regular users will have to wait for the broader rollout.