Google launches Gemini 4 Argon — but you can't call it yet
Google's new frontier model matches GPT-6 Astra at ~60% the cost with 1M-token output — but ships only to vetted cyber defenders, and trails on coding.
Copy markdown
A top-tier model, locked behind Fairwind
Gemini 4 Argon shipped Oct 1 but there's no public endpoint yet — access runs through Google's Fairwind cyber-defender program and a US-government pre-release, with no model ID live in AI Studio, Vertex, or the API. When it opens, intro pricing is $2/$10 per million tokens (standard $4/$20, cached input discounted ~95%), and max output jumps to 1M tokens from 64K — a single call can emit a whole codebase or book. For now, don't plan a launch around it; you can't call it today.
Strong on knowledge, soft on coding
Argon tops Text Arena (1525) and scores 53 on Artificial Analysis's Intelligence Index — matching GPT-6 Astra at roughly 60% of the cost per task. But it trails exactly where agent builders live: FrontierSWE v2 at 55.0% (GPT-6 Astra 65.5%) and Terminal-bench 4.0 at 57.4% (Claude Opus 5.5 66.4%). Great for long-context knowledge work; not an obvious upgrade for your coding agent.
Why it's gated: cyber guardrails
Google says Argon ships with 'misalignment mitigations' that monitor its chain-of-thought and halt execution mid-run, and claims it leads Gray Swan's indirect-prompt-injection benchmark. A guardrail-free build is reserved for trusted defenders and Google's own teams. Expect the 'frontier capability gated behind a vetting program' pattern to spread as models get more autonomous — plan for staged access, not day-one availability.
DoGBench: no model writes your docs
A new open benchmark, DoGBench, tests user-facing documentation generation — and no frontier model clears 50%. If you've wired an LLM into a docs pipeline, treat its output as a draft that needs human review, not a shippable page. A concrete eval you can point your own setup at.