Google DeepMind 推出 Gemini 4 Argon，產量達 100 萬枚並具有新定價

發生了什麼事

Google DeepMind unveiled Gemini 4 Argon, the inaugural model of the Gemini 4 generation, highlighting a dramatic increase in output length—from 64 K tokens on prior Gemini models to a full 1 million tokens in a single response. The model is positioned for long‑horizon software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense. Pricing is publicly disclosed: an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted to $0.10 per million. After the introductory period, rates rise to $4 input and $20 output. Benchmarks show Argon leading on 12 of 18 tests and tying for first on one, and on the CWE‑bench v1 vulnerability‑remediation suite it ties for first with a 68 % score. Early adopters such as Wiz’s Scan for Good initiative report that Argon identified a critical vulnerability missed by earlier frontier models. Google is rolling out the model through a phased, voluntary pre‑release program for U.S. government testers and will refine guardrails before broader availability.

Google DeepMind announced Gemini 4 Argon on September 30, 2026, describing it as the first model of the Gemini 4 generation. The headline technical advance is a 1 million‑token output limit, a twenty‑five‑fold increase over the 64 K token ceiling of earlier Gemini models.

Pricing is set at $2 per million input tokens and $10 per million output tokens for the introductory period, with cached input tokens discounted to $0.10 per million. After the introductory phase, rates double to $4 input and $20 output. Logan Kilpatrick, a Google spokesperson, confirmed these figures.

Benchmark testing released by Google shows Argon leading on 12 of 18 standard AI benchmarks and tying for first on one. On the CWE‑bench v1 vulnerability‑remediation test, Argon ties for first with a 68 % success rate, outperforming rival models that run inside separate agent harnesses.

Early adopters such as Wiz’s Scan for Good initiative report that Argon discovered a critical vulnerability in widely used healthcare software that prior frontier models missed, illustrating the model’s potential in high‑impact security contexts.

Google is conducting a phased rollout, participating in the U.S. government’s voluntary pre‑release access program to gather feedback and refine safety guardrails before a broader commercial launch.

來源詳情: marktechpost.com ↗

為什麼這很重要

The 1 million‑token context window unlocks use cases that previously required splitting work across multiple API calls, such as massive code refactors, exhaustive legal briefs, or comprehensive security analyses. By offering a pricing structure that heavily discounts cached inputs, Google signals an intent to make large‑scale, repetitive workloads economically viable, though the $10‑$20 cost for a full‑length output remains a tangible expense for many developers. Benchmark dominance suggests Argon could become the de‑facto tool for high‑stakes enterprise and security tasks, potentially shifting competitive dynamics with Anthropic’s Claude series and OpenAI’s GPT‑6 Astra. The model’s early deployment in vulnerability‑remediation workflows also raises questions about the balance between powerful AI assistance and the need for robust safety guardrails, especially as Google plans to extend access beyond trusted defenders.

The expanded token window enables developers to generate extensive codebases, long‑form reports, or detailed security analyses without breaking the task into multiple API calls, reducing latency and simplifying workflow orchestration.

Pricing that heavily discounts cached inputs suggests Google is targeting repetitive, high‑volume enterprise workloads, but the $10‑$20 cost for a full‑length output remains a barrier for smaller teams or hobbyist developers.

Benchmark superiority positions Argon as a strong contender in the competitive landscape, potentially prompting Anthropic and OpenAI to accelerate their own context‑length or pricing strategies.

Deploying Argon in vulnerability‑remediation workflows highlights both the promise of AI‑driven cyber defense and the necessity of robust guardrails to prevent misuse or over‑reliance on automated patching.

接下來看什麼

Key indicators to monitor include the rollout timeline of the phased pre‑release program, any adjustments to the introductory pricing after the launch period, and the evolution of Google’s guardrails for cybersecurity use. Competitor responses—particularly pricing or context‑window changes from Anthropic and OpenAI—will reveal how the market values extreme token lengths. Adoption metrics from early partners like Wiz will indicate real‑world cost‑benefit outcomes, while any reported misuse or safety incidents could prompt regulatory scrutiny or policy interventions.

The schedule and scope of the phased pre‑release program, especially which sectors receive early access and how feedback shapes model safeguards.

Any revisions to the introductory pricing model after the launch period, which could affect adoption rates among cost‑sensitive enterprises.

Responses from competing AI providers—whether they increase token limits, adjust pricing, or introduce new safety features—to maintain market parity.

Real‑world performance data from early adopters like Wiz, including cost‑per‑vulnerability metrics and any reported safety incidents that could trigger regulatory attention.

相關指引和測驗