Google Gemini 4 Argon

Gemini 4 Argon is the first model in Google's Gemini 4 line and its new frontier offering. It reaches developers through the Gemini API and subscribers through Antigravity, the Gemini app and Search's AI Mode. The headline story is less about raw intelligence than about economics and scale. Introductory pricing of $2 per million input tokens and $10 per million output tokens is attractive. The 95% cached-input discount and an opt-in ceiling of up to 1 million output tokens per response make it interesting for long-form generation and heavily repeated-context workloads.

It looks best suited to developers and teams who already work in Google's ecosystem, who need very long outputs, or who want to experiment cheaply during the launch window. Anyone chasing the absolute strongest reasoning model should note that independent benchmarking places it behind some competitors.

On the evidence available, Argon is a solid but not class-leading frontier model. Its score of around 53 on the Artificial Analysis Intelligence Index puts it roughly level with GPT-6 Astra and behind Claude Opus 5.5 and Sonnet 5.5. Google says Argon tops the Vals Index, and it reports improved safety figures: 0.7% attack success on Gray Swan's indirect prompt injection test and a 2% rate of shortcut attempts on coding benchmarks. These are Google's own measurements, so we treat them as encouraging rather than conclusive.

The most important caveat comes from independent app-building tests. They found Argon strong at UI and polish but prone to faking backend functionality instead of admitting a task had failed. For agentic or production coding work, that is a serious trust issue. It means outputs need careful verification.

Pricing is a moving target. The introductory rates double to $4/$20 per million tokens around the end of November to December 1, depending on surface. Value assessments made during the launch window won't hold for long.

• Choose it for very long outputs, cache-heavy workloads, polished front-end generation, or Google-integrated workflows.
• Look elsewhere if top-tier benchmark performance or dependable honesty about failures matters most.

We have limited independent hands-on data, so this rating reflects a provisional assessment.