Google announced Gemini 4 Argon on September 30, calling it its new frontier model for real-world software engineering, enterprise knowledge work, and cybersecurity defense. But unlike every other flagship launch this year, you can't use it: Argon is rolling out first to a small set of trusted cyber defenders through something called the Fairwind Program, with paid API access and AI Ultra coming later. After covering the GPT-5.6 government-review saga and the GPT-6 Astra launch from this same desk, here's my honest take: the model is impressive, but the rollout is the real story.
This guide explains what Argon actually is, which benchmark numbers matter, who gets access and when, what it costs (spoiler: Google hasn't said), and – the question I care about most – what a defender-first flagship means for people running local AI on affordable hardware.
Source note: specs and scores below come from Google's launch announcement and independent reporting linked at the bottom. There is no public API pricing yet, so Argon is not in our AI API cost calculator – I'll add it the day pricing goes public.
The short version
- What: Gemini 4 Argon, Google's new frontier model, built for deep reasoning across long, multi-step workflows.
- Standout spec: 1 million token output limit, up from 64,000 – roughly 15× more continuous generation than before.
- Scores: 77.9% on DeepSWE v1.1 (agentic coding), 68% on CWE-bench v1 (vulnerability work), 91.7% on LVBench (long video understanding).
- Access: trusted cyber defenders first (Fairwind Program), then paid API customers and AI Ultra subscribers. No dates for either next step.
- Price: unpublished. No per-token rates, no calculator entry yet.
- Safety: phased release under the US government's voluntary pre-release process – the same playbook as GPT-5.6's limited preview.
What Gemini 4 Argon is
Argon is the first model of the Gemini 4 generation, and Google is positioning it less as a chatbot brain and more as a coworker for work that takes hours: refactoring a large codebase, drafting and reviewing legal and finance documents, and – the headline use case – finding and autonomously patching software vulnerabilities. The through-line in Google's announcement is long-horizon reliability: not answering one hard question, but staying coherent across hundreds of steps.
The 1M-token output limit is the number I'd circle. Input context gets the headlines, but output limits are what actually cap agent runs today: at 64K output tokens, a coding agent has to stop, summarize, and restart every hour or so, losing state each time. A million output tokens means an agent can, in principle, work a full shift without amnesia. Whether quality holds up across that much generation is the open question no benchmark fully answers yet.
The benchmark numbers that matter
Google reported four headline scores. As always, vendor benchmarks are the start of the conversation, not the end – but two of these are unusually informative:
- DeepSWE v1.1 – 77.9%: multi-step software engineering tasks. This is the score most correlated with "can it actually do my ticket", and 77.9% is frontier territory.
- CWE-bench v1 – 68%: real-world vulnerability detection and patching across Common Weakness Enumeration categories. This justifies the defender-first rollout: the capability and the risk are the same thing.
- LVBench – 91.7%: long-video understanding. Relevant if you work with meeting recordings, lectures, or surveillance footage – and a sign the long-context story extends beyond text.
- AutomationBench – 51.3%: autonomous multi-step task completion. The lowest of the four, and arguably the most honest: we're still at coin-flip reliability for fully unsupervised work.
Independent confirmation will take weeks – Argon needs to reach third-party evaluators first. Until then, treat the ranking as "frontier-class" rather than "best".
Access: Fairwind first, everyone else later
The rollout order tells you what Google is worried about. Phase one is the Fairwind Program: a set of trusted cyber defenders – think enterprise security teams, not the public. Phase two is paid API customers and AI Ultra subscribers. No dates, no waitlist, no self-serve signup as of this writing.
This is now a pattern, not an exception. GPT-5.6 shipped as a limited preview under US government review; Argon ships to defenders under a voluntary pre-release process. Frontier releases are converging on the same shape: capability review first, broad availability later, with the most security-sensitive capabilities gated longest. If you build on top of frontier APIs, plan your roadmaps around phased access, not launch-day availability.
Cybersecurity: the capability is the risk
A model that scores 68% on CWE-bench can find and patch vulnerabilities at machine speed – in your codebase, and in everyone else's. Google's answer is to give it to defenders first, which is reasonable triage but worth stating plainly: every month of defender-only access is also a month where the same capability class exists but isn't broadly available for defense. The bottleneck in vulnerability management is already moving from finding flaws to fixing them faster than attackers can exploit them, and Argon-class models accelerate both sides.
Price and availability
Unpublished. Google has not released per-token API rates, batch discounts, or Ultra plan changes tied to Argon. That means no entry in our cost calculator yet – every number there is verified against an official pricing page, and I'm not going to guess. One thing to watch when pricing does land: long-output models shift spend from input to output tokens, so compare cost per finished task, not cost per million tokens.
What this means if you run local AI
Three honest implications from someone who benchmarks open models on an RX 6800M every week:
- The long-horizon gap widens. Local models are catching up on single-answer quality, but 1M-token coherent output with 77.9% DeepSWE reliability is not something a 32B quant does today. For multi-hour autonomous coding, the cloud frontier just pulled away.
- Defenders get leverage first. If you self-host security tooling, watch what Fairwind participants publish – the agentic-vuln-patching workflow they develop will eventually be replicable with open models, just later and cheaper.
- The local case still holds for everything else. Privacy, fixed cost, offline use, and short-form tasks haven't moved. Argon changes nothing about the economics of running a 27B model for drafting, summarization, and RAG on your own box.
Bottom line
Gemini 4 Argon looks like a genuine frontier step – the 1M output limit and the SWE/vuln scores are the substance – wrapped in the most cautious rollout of any flagship so far. Google is treating its best model like infrastructure with a blast radius, not a product with a launch date. For most of us, the practical effect is: read the benchmarks, wait for API access, and keep an eye on pricing day, when I'll update the calculator. If you're a defender in the Fairwind Program, the rest of us would love to hear what it actually does – my inbox is on the contact page.