Google DeepMind has officially unveiled Gemini 4 Argon, marking the debut of its fourth-generation model family. The company is placing a massive expansion of output capacity at the center of the release, introducing an unprecedented output limit of one million tokens per single response. Typical frontier models have previously capped generation at 64,000 or 128,000 tokens, with competing systems such as GPT-6 Astra reaching approximately 262,000 tokens. This expanded generation ceiling is specifically designed to support prolonged reasoning workflows and extensive test-time compute.
Google backs the architectural changes with concrete benchmark results, particularly in autonomous software engineering. On the rigorous DeepSWE benchmark, Gemini 4 Argon achieved a score of 77.9 percent, representing a substantial gain over the approximately 65 percent recorded by Gemini 3.7 Flash. On AutomationBench, which evaluates multi-step automation routines, the new model reached 51.3 percent. These results underscore an industry-wide pivot toward models capable of tackling end-to-end coding tasks without human guidance.
A core application area for Argon is cyber defense and automated code remediation. Testing on CWE-bench v1 shows the model achieving a 68 percent resolution rate, tying with GPT-6 Astra. Google is using this foundation to deploy CodeMender, an automated vulnerability patching system designed to independently identify and remediate security flaws across more than twenty programming languages.
Despite its technical milestones, Google faces widespread pushback over its distribution strategy. Access to Gemini 4 Argon is strictly restricted, initially available only through the Fairwind safety evaluation initiative and the newly introduced Google AI Ultra subscription tier. This new enterprise tier is priced between 100 and 200 dollars per month, presenting a significant barrier to entry for individual developers and smaller teams.
The pricing structure has sparked intense frustration among existing Google Pro subscribers, who find themselves entirely shut out from Argon despite their ongoing monthly commitments. The exclusive rollout highlights the economic realities of operating cutting-edge reasoning models, as tech companies increasingly segment frontier capabilities into premium price tiers to offset escalating compute expenses.