Google has unveiled Gemini 4 Argon, its new flagship model, but is initially giving access only to selected cybersecurity partners and a US government pre-release programme.
The restricted debut matters as much as the model itself. Google says Argon is its largest and most capable Gemini model for complex work, yet developers have no public release date and cannot independently test the benchmarks used to position it against OpenAI and Anthropic.
Why it matters
A security engineer choosing a model for vulnerability work cannot buy Argon today, reproduce Google’s scores or compare its safeguards under ordinary deployment conditions. The launch instead gives a small group an early look at a system Google describes as especially capable in cyber tasks.
Google’s announcement presents Argon as the first model in the Gemini 4 generation. According to Reuters, the company says the model is larger than its previous Pro line and comparable with leading rivals on selected coding and cybersecurity tests. Google’s own tables put Argon ahead on some measures but behind on two of the four coding benchmarks it disclosed.
That mixed result is more useful than a clean sweep would have been. It shows that “frontier” still describes a bundle of strengths, not a single finish line. A model can lead a cyber test while trailing on parts of software engineering, and the choice of benchmark determines which story gets told.
Google has not disclosed a public ship date. It is providing Argon to vetted cyber defenders and participating in the US government’s voluntary process for giving officials access before release. Reuters also reported that Google abandoned Gemini 3.5 Pro, which chief executive Sundar Pichai had previously said would arrive in June.
The cancellation helps explain why Argon is being introduced as both a technical release and a reset. Google spent much of 2026 emphasizing smaller, cheaper models while Anthropic and OpenAI refreshed their top tiers. Argon gives Google a new flagship name, but the limited-access phase postpones the market test that matters: whether the model’s capability, latency and price remain competitive outside Google’s controlled evaluations.
That delay also changes the buying decision. Teams can note Google’s benchmark claims now, but they cannot yet measure throughput, tool reliability or total task cost in their own workloads. Those deployment results, rather than the launch label, will determine whether switching models is worthwhile.
Independent coverage adds an important boundary to the announcement. Reuters confirmed the limited availability and the abandoned Gemini 3.5 Pro plan through a company spokesperson, while also noting that Google supplied the performance numbers. The Financial Times reported an initial price of $2 per million input tokens and $10 per million output tokens, but public access and final commercial terms remain unsettled.
Google’s cautious rollout is defensible for a model aimed at cyber work, where a capability gain can help defenders and attackers. It also concentrates evidence in the hands of the vendor and its chosen partners. The two facts are inseparable.
The next concrete milestone is broader developer access. Until Google names that date and publishes stable commercial terms, Argon is a flagship announcement with a controlled evaluation audience, not a generally available replacement for the models teams can deploy now.