Google introduces Gemini 4 Argon but holds back broad access

The story

Google has introduced Gemini 4 Argon, a new frontier model designed for long-running work in software engineering, professional research and cybersecurity. Unlike a conventional product launch, however, Argon is not yet broadly available. Google is first giving access to a selected group of cyber defenders through its Fairwind Program while it gathers feedback and strengthens safeguards before opening the system to developers, enterprises and consumers.

That controlled release is central to the announcement. Google says it is also participating in the U.S. government's voluntary process for pre-release model access. The company has not provided a date for general availability, telling users only that paid API customers and Google AI Ultra subscribers will be first in line when access expands. Reuters independently confirmed the restricted rollout and the absence of a public-release timetable.

Argon anchors the Gemini 4 generation and replaces the previously expected Gemini 3.5 Pro at the top of Google's model line. Reuters reported that Google no longer plans to release 3.5 Pro, which had originally been expected in June. The reset arrives after months in which Anthropic and OpenAI advanced their own leading systems, raising pressure on Google to show that its next flagship could compete on complex coding and agentic tasks.

Google's technical pitch emphasizes endurance. Argon supports an output limit of one million tokens, up from 64,000 in the previous generation, according to the company. That is an output ceiling, not proof that the model can remain accurate through every extremely long task. The intended advantage is enough working room for extended software migrations, research sequences and other multi-step trajectories that would otherwise be split across many runs.

The company reports a score of 77.9 percent on DeepSWE v1.1 for long-horizon software engineering, first place on the Vals Index for economically weighted professional tasks and 51.3 percent on Zapier's AutomationBench. Google also says Argon ties for the top score on CWE-bench v1 for repairing software vulnerabilities. These results are useful signals, but they should be treated as vendor-selected evidence until independent researchers can reproduce them under comparable conditions.

Reuters added an important qualification: although Google's release showed Argon ahead of rival models on several self-reported measures, it remained behind on other tests, including two of the four coding benchmarks included by Google. Benchmark results are also sensitive to prompting, tool access, test contamination and the amount of computation allowed. No single score establishes reliability across production environments.

Google says it has already used Argon internally for large code migrations, quantum-algorithm optimization and data-centre efficiency. One internal deployment reportedly found memory optimizations that freed more than 300 tebibytes after rollout, with additional savings estimated. Another project used agents to help migrate C and C++ code to Rust. Those examples have not yet been independently audited, and Google notes that critical code changes still undergo automated and human review.

Cybersecurity is where access policy becomes unusually nuanced. Google says trusted defenders and its internal teams will receive the model without the cyber restrictions applied to general users so they can investigate and patch serious vulnerabilities. For wider use, the company is working on controls against cyber and chemical, biological, radiological and nuclear misuse, indirect prompt injection and agents acting beyond a user's intent. It also says it monitors internal reasoning and actions to halt unsafe execution when necessary.

The commercial strategy is aggressive. Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 percent. After the introductory period, those rates are scheduled to double to $4 and $20. Pricing will matter as much as peak capability for enterprises running lengthy agent workflows, because a model that produces hundreds of thousands of tokens can create substantial inference costs even at a low per-token rate.

INNOVOX analysis: Argon's most consequential feature may be the gap between announcement and access. Google is asking the market to evaluate a frontier system largely through controlled partners, internal deployments and selected benchmarks while acknowledging that safety work is unfinished. A staged release can reduce risk and improve defenses, but it can also delay independent scrutiny. The next test is therefore not another chart. It is whether outside users can reproduce the claimed gains without new security failures, runaway costs or a level of human supervision that erases the promised productivity advantage.

INNOVOX analysis

The significant change is not one benchmark score but the release model. Google is separating a high-capability system from immediate mass distribution, using specialist cyber partners and government pre-release access to test safeguards first. That creates a potentially useful precedent, but its credibility will depend on what independent evaluators learn and how transparently Google reports failures, not only successful demonstrations.

What to watch

Watch for an independent system card, third-party benchmark replication, a firm public-release date and details of which capabilities remain restricted. The most revealing evidence will be real-world task completion, security incidents, refusal reliability and operating cost after customers can test the model outside Google's controlled environment.