Quick Facts
- Google made Gemini 4 Argon generally available on Oct. 13, 2026, two weeks after a limited cybersecurity-focused launch on Sept. 30.
- Introductory API pricing is $2 per million input tokens and $10 per million output tokens through Nov. 30, after which rates double to $4 and $20.
- Argon leads or ties on 13 of 18 disclosed benchmarks, outperforming Claude Opus 5.5 and GPT-6 Astra in legal, finance, automation, and long-context tasks.
Google opened Gemini 4 Argon to all developers and enterprises on Oct. 13, 2026, making its flagship reasoning model available through the Gemini API, Google AI Studio, and its Antigravity agentic development environment. Google AI Pro and Ultra subscribers can also access the model directly in the Gemini app.
The model had been restricted to cybersecurity partners through Google's Fairwind Program since its Sept. 30 debut. That program gave governments, healthcare providers, and telecommunications companies early access to the model ahead of general availability.
Pricing Window Closes Nov. 30
Google set a hard deadline on its introductory rates. Standard API access runs $2 per million input tokens and $10 per million output tokens through Nov. 30. On Dec. 1, those prices double to $4 and $20 per million tokens.
Batch and flexible tier pricing runs at $1 and $5 per million tokens during the promotional period, rising to $2 and $10 after the deadline. Priority tier access costs $3.60 and $18 now, scaling to $7.20 and $36 in December. Cached input carries a 95% discount across all tiers.
At current introductory rates, Argon costs roughly one-fifth what Anthropic charges for Claude Fable 5.1 and what OpenAI charges for GPT-6 Astra, both priced at $10 and $50 per million tokens. After Nov. 30, Argon's standard pricing matches Anthropic's Claude Opus 5.5 at $4 and $20 per million tokens.
Architecture and Output Limits
Argon is a sparse mixture-of-experts model built for multi-step tasks across coding, finance, legal work, science, and cybersecurity. Google raised the output ceiling to 1 million tokens, up sharply from the 64,000-token limit on earlier Gemini models.
Default replies cap at 65,536 tokens. Background requests can run to approximately 1 million tokens minus input. The combined input and output window also tops out at 1 million tokens.
Benchmark Performance
Argon leads or ties on 13 of 18 disclosed benchmark categories against Claude Opus 5.5 and GPT-6 Astra. It scores 77.9% on DeepSWE v1.1, compared to 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. On Harvey's Legal Agent benchmark, Argon scores 19.6%, versus 3.8% for Opus 5.5 and 5.4% for Astra.
Argon ranks first on Zapier's AutomationBench at 51.3%, a test measuring end-to-end execution across core business functions. It also hits 65.4% on Vals Finance Agent v2, ahead of Claude Fable 5.1 at 58.9% and GPT-6 Astra at 53.5%.
The model trails competitors in terminal-based and large-codebase tasks. GPT-6 Astra holds a 10.5-point lead on FrontierSWE v2, and Claude Opus 5.5 leads by 9 points on Terminal-Bench 4.0.
Security Capabilities
Google reports Argon achieved a 0.7% attack success rate on Gray Swan's 2026 Indirect Prompt Injection benchmark, the lowest score among 13 models tested. The model also scored 68% on CWE-bench v1, a vulnerability repair test, tying for the top position.
Cloud security firm Wiz used Argon during the Fairwind early-access period to identify vulnerabilities in government and healthcare infrastructure. One assessment uncovered a flaw that exposed a master key capable of granting administrator access to healthcare systems.
Argon is Google's first new flagship model generation since Gemini 3 shipped last November. Teams requiring the model without its built-in cybersecurity guardrails can still apply through the Fairwind Program.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.