Gemini 4 Argon launches at $2 per million input tokens

Google is pitching its newest frontier model on price before it has finished releasing it. Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens during an introductory period, with cached input 95% off. After that, the rate doubles to $4 and $20.

Google DeepMind announced Argon on 30 September 2026. It is the company's first new frontier model generation since Gemini 3, aimed at long-horizon coding, legal and finance work, and cyber defence. Google positions it against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5.

Access is staged. Trusted cyber defenders in Google's Fairwind Program get the model first, along with US government pre-release access. Paid API users and Google AI Ultra subscribers follow. Google says broader developer, enterprise and consumer release is "coming soon", without giving a date.

Google also raised the output limit to 1 million tokens, up from 64K, and describes that as industry-leading. The company's own benchmark figures lead with coding and security:

Google also says Argon leads the Vals Index across finance, coding, legal and tax, and leads Gray Swan's prompt-injection robustness tests. These are the company's figures, not independent results.

Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, offered an internal use case in the launch post: "Argon is helping our quantum computing researchers optimize the spacetime resources (qubits × gates)".

Outside reaction is guarded. CNBC asked whether Argon can catch OpenAI and Anthropic at the frontier. One analyst told it that "the final proof will be in enterprise production environments once the model is fully released". Bloomberg reported that some Google employees privately doubt the model's real-world coding ability. Google has not commented on that report in the notes available, and it remains unconfirmed.

The pricing structure is the story. An introductory rate that doubles gives Google a window to win trial workloads and test production use while buyers are still comparing. Teams that build on the $2 and $10 rates should budget for $4 and $20, because the cheaper figure is temporary.

Google is leaning on the segments where it can show numbers: coding, vulnerability remediation, finance, legal and tax. Those are also the workloads where enterprises pay most and switch least readily. The cyber-defender-first rollout fits that pitch, since it puts the model in front of security buyers before the wider market.

The exposure is the gap between benchmarks and production. Every headline score is Google's own, the analyst's test is enterprise deployment, and the Bloomberg report targets coding, the area Google leads with. Watch the general release date, what happens to demand when the price doubles, and whether independent results back the DeepSWE and CWE-bench claims.