Google has announced Gemini 4 Argon, the first model in the Gemini 4 family. It is a vision-language model built for long, complex work in software engineering, law, finance, and cybersecurity. It accepts text, images, and video (up to 1 million tokens) and can output up to 1 million tokens, up from 64,000 for Gemini 3.1 Pro Preview. That is made possible by a Gemini API feature called Long Decode Continuation, which pauses long responses and resumes them in later calls. For now, only organizations in Google’s Fairwind cybersecurity program have access. Their API introductory price is $2/$0.10/$10 per million input/cached/output tokens, rising later to $4/$20 per million input/output tokens. The model is proprietary, and Google has not disclosed its size, architecture, or training details.
In independent benchmarks, Argon ranks first on Vals AI’s Vals Index (68.9 percent) and on Arena.ai’s Text Arena (1,525 Elo). It ties for third on Artificial Analysis’ Intelligence Index with a score of 53, matching GPT-6 Astra and Claude Fable 5.1 at lower cost per task. It has the lowest hallucination rate (15 percent) of any model scoring 45 or higher on that index, versus 51 percent for GPT-6 Astra. It still trails Anthropic’s newest models on agentic coding. Google follows Anthropic and OpenAI in gating a powerful model: selected defenders get a version without cybersecurity guardrails, while the wider release will refuse harmful cyber, chemical, biological, radiological, or nuclear requests. Safeguards include probes that monitor internal activations and training against prompt injection.
Why it matters
- More trustworthy answers: Argon is described as the frontier model most likely to admit it doesn’t know a factual answer rather than guess wrong, which matters for finance, tax, and legal work where errors are costly.
- Not widely available yet: Only Fairwind cybersecurity organizations can use it, and Google hasn’t given a date for broader access or an end date for introductory pricing, so most businesses will need to wait.
- Security payoff: Google says its agents translated C and C++ code into Rust, including more than 800,000 lines from the Fuchsia kernel. If this works cheaply, defenders could remove whole categories of memory bugs rather than patching them one by one.
Source
- Google’s Gemini 4 Argon: first Gemini 4 model, lowest hallucination rate, leads finance/legal benchmarks (limited access for now) — The Batch, 10 Oct 2026
Summary written by FoxaMind with AI assistance from the source above. Check the original for full details.