Google Launches Gemini 4 Argon for Initial Network Defense Trials
Google has released its flagship model Gemini 4 Argon, initially available to trusted network defense parties under the Fairwind program, and not yet open to general developers and consumers.
The company states that this model is designed for long-cycle complex workflows, covering enterprise knowledge work in software engineering, law, finance, and cybersecurity defense. The output limit has increased from 64,000 tokens to 1 million tokens. DeepSWE v1.1 scored 77.9%, higher than GPT-6 Astra's 74.1% and Claude Opus 5.5's 74.2%. The Vals index leads, measuring financial, coding, legal, and tax tasks by U.S. GDP weight. LVBench long video understanding scored 91.7%. CWE vulnerability remediation benchmarks are tied for the lead with Astra. Internal vulnerability discovery scored 85.8%, and Wiz penetration testing scored 70.9%, both higher than 3.8 Flash Cyber.
The introductory price is $2 per million input tokens and $10 for output, with a 95% discount on caching input relative to input pricing; after the introductory period, prices will rise to $4 and $20. The next step is to provide access to paid interface users and subscribers of Google AI Ultra at $200 per month. Google has not provided a timeline for full access, stating it is participating in the U.S. government's voluntary pre-release model access assessment. Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at DeepMind, stated that access will be gradually expanded. Product lead Tulsee Doshi mentioned that the initial focus on defense parties is to address vulnerabilities found in the model before wider release. The model includes out-of-bounds monitoring, which can cut off activity when deviating from user intent.
The 3.5 Pro was originally scheduled for release in June but was delayed multiple times due to poor internal performance and was not released, with a smaller Flash being the main push in the summer. Argon is larger than the previous Pro line. Internally, Google has used it for specialized coding, research, and writing; agents are assisting in migrating C/C++ to Rust, including over 800,000 lines of re2, libgav1, and the Zircon kernel of Fuchsia; the video decoder is 2.7 times faster than the Rust port and produces consistent output. The telemetry from the cluster reportedly saves about 300 terabytes of memory. The day before the release, Sundar Pichai and other tech leaders signed a voluntary security commitment at the White House.
Computing power and interface quotas are initially directed to government and defense agencies, rather than consumer chat windows. Buyers are network defense parties needing to patch first and enterprises requiring long output coding agents, while the seller is Google, which has changed its release rhythm to prioritize security reviews. This event is driven by benchmark battles against Astra and Opus, while using phased releases to respond to alignment controversies. Beneficiaries are the defense teams within Fairwind and Google's internal migration projects; those under pressure are developers waiting for public interfaces and customers who were disappointed by the 3.5 Pro expectations. The 1 million output extends the single inference trajectory, aligning the latter half of the price war with Opus output pricing.
Source: Public Information
ABAB AI Insight
After Gemini 3, Google has eliminated the 3.5 Pro, filled the gap with Flash, and then jumped to the fourth generation with Argon. The path is to first feed the model to internal code migration and memory-saving in data centers, then to Fairwind defense parties, and finally to Ultra subscribers and paid interfaces. This is in contrast to OpenAI and Anthropic, which prioritize consumer access before discussing security: first proving the ability to find vulnerabilities and modify kernels, then addressing chat windows. Pichai signed at the White House the day before, and the next day completed a political and product synchronization by stating "only for defense parties."
Capital is shifting from training clusters to billable million-level outputs. The introductory price of $2/$10 targets coding agent budgets, with a 95% discount on caching reducing long context repetition costs, and prices doubling after the introductory period to align with Opus output pricing. Money has not yet come from public subscriptions, but is being realized from the 300 terabytes of memory saved internally and the Rust migration. The motivation is to frame "cutting-edge models are too dangerous" as part of the release process rather than halting training. Resources are prioritized for institutions that can verify vulnerabilities, effectively exchanging government access for a time window.
In contrast to Microsoft embedding models into GitHub and defense clouds, Anthropic trading constitutional and refusal strategies for enterprise contracts, and OpenAI using Astra to benchmark coding. Google is in a control period to rebuild the cutting-edge narrative with the initial focus on network defense: after the failure of 3.5 Pro, it must hit DeepSWE and Vals in one go, or the consumer side will continue to lose. The industry is shifting from comparing parameters and chat fun to comparing who can complete migration and patching within a single 1 million output.
The essence is a change in regulation layered with technological substitution. Long-cycle agents that can modify kernels and find vulnerabilities shift the release authority from product managers to defense reviews. The mechanism is: the closer the capability is to the attack surface, the more it must be given first to the defense side; whoever controls the pre-release access list decides how long opponents must wait to use the same tool. Leading benchmarks only open the negotiation table; Fairwind decides when the knife is drawn.
ABAB News · Cognitive Laws
Models that can modify kernels must first be given to those fixing vulnerabilities.
Extending output limits allows a single inference to compensate for migration.
The release list is closer to power than parameter scale.