On September 30 we wrote that Gemini 4 Argon was phased for a reason. Access started with cyber defenders in Google’s Fairwind Program, and most teams couldn’t touch it yet.
That changed on October 13. Here’s what’s new, and what it means if you build agents for a living.
It’s available, with a real price clock
Argon is now in the paid tier of the Gemini API and AI Studio, in Gemini for business, and in the Gemini app for Google AI Ultra and Pro subscribers. It isn’t on the free API tier.
Google has also given the introductory price an end date. It’s $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1. Cached input is 95% off. If you’re budgeting a pilot, plan for the second number, not the first.
The 1M token limit is shared
On September 30 we reported a 1M token output limit. The developer docs are more specific. Input, thinking and output share one 1,048,576 token budget. Synchronous calls still default to 65,536 output tokens, and longer runs use a new continuation protocol. Thinking tokens bill as output.
For agent builders, that’s a design question, not just a bigger number. A long run can spend its budget on thinking you never see, so cap it and log it.
Independent numbers are in
Two weeks ago, almost every number was Google’s. Now there’s outside data. Artificial Analysis puts Argon at 53 on its Intelligence Index, tied with GPT-6 Astra and Claude Fable 5.1 and behind Claude Opus 5.5 at 58. Reviewers note it got there at a lower cost per task than Astra.
Google’s own table still shows Argon leading most rows, including #1 on the Vals Index, while trailing on five, including Terminal-Bench Science and OSWorld-2.0. Those are Google’s claims. Our read hasn’t changed: it’s a strong option, and the right model depends on the workload.
Google is now making agent safety claims
This is the part we’d point builders to first. Google says it trained Argon to respect system boundaries, report negative findings honestly, and pause for human confirmation before irreversible steps. It reports shortcut attempts on coding benchmarks fell to 2%, and attack success on the Gray Swan prompt injection benchmark fell to 0.7%. It also describes a monitor that reads the agent’s reasoning and actions to block workarounds.
Those are good defaults, and they’re Google’s claims. A model that pauses on its own is helpful. It isn’t a substitute for an approval step you designed, a log you can audit, and permissions that make the dangerous action impossible in the first place.
What we’d do this week
- Test it on a workflow you already run. Same tools, same approval gate, same inputs. Compare cost, retries and failures, not demo quality.
- Budget for December prices. Put a token cap on every run.
- Keep the gate. Irreversible writes still go through a human.
Google’s Fairwind Program is also expanding to hundreds more defender partners, so the cyber version still sits behind a separate gate.
If you want help putting a new model into a bounded workflow, with evidence and a human in the loop, that’s what we build at Viking Labs.
Sources:
Leave a Reply