Gemini 4 Argon arrives for cyber defenders first, with a 1M-token output budget and a $2/$10 intro price

Google announced Gemini 4 Argon on Wednesday, calling it its new frontier model. The news is less a launch than a staged opening. Per Google's announcement, Argon is rolling out first to a group of cyber defenders in its Fairwind Program. Developers, enterprises and consumers come later, with no date given. Google says it is taking part in the U.S. government's voluntary pre-release access process while it widens access and keeps iterating on guardrails.

What developers can plan around now is the price, the output budget and the order of the queue. The benchmark claims are Google's own and have not yet been independently reproduced.

What Google actually announced

• Access: the first users are trusted cyber defenders in the Fairwind Program. Google says the broader release will start with paid API customers and Google AI Ultra subscribers, but gives no date and has not published a public model ID.
• Price: an introductory $2 per million input tokens and $10 per million output tokens, with cached input at 95% off the input price. A footnote says the price rises to $4 input and $20 output when the introductory period ends. Google does not say when that is.
• Output limit: 1 million output tokens per response, up from 64K on earlier Gemini models. Google calls that industry-leading.
• Positioning: long-horizon software engineering, enterprise knowledge work such as legal and finance, and defensive cybersecurity.

The Fairwind Program dates from September 2, when Google opened it to governments, critical-infrastructure operators and core technology platforms. Until now it gave them access to Gemini 3.8 Flash Cyber and the CodeMender remediation harness. Google says trusted defenders and its own internal teams will get Argon without cyber guardrails, so they can use its full vulnerability-finding capability. It is unusual for a lab to say that so plainly, and it explains why access is restricted.

The numbers are Google's numbers

Google says Argon scores 77.9% on DeepSWE v1.1, a long-horizon software engineering benchmark, and calls that a new state of the art. It also gives 51.3% on Zapier's AutomationBench (ranked first), 91.7% on LVBench for long-video understanding, and 68% on CWE-bench v1, tied for first on vulnerability remediation. Google also says Argon leads the Vals Index, a composite of finance, coding, legal and tax work weighted by each sector's share of U.S. GDP.

All of these come from the announcement. Benchmarks like these are useful as a direction, but Google picked which ones to show. TechCrunch points out that the blog compares Argon favorably with OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models, and describes that as part of the labs' marketing race. Until Argon is generally available and third parties can run it, treat these figures as vendor claims, not rankings.

The internal examples deserve the same caution, though they are more concrete than most launch anecdotes. Google says Argon agents freed more than 300 TiB of memory across its data centers through fleet-wide profiling. It says they are migrating C and C++ codebases to Rust, including more than 800,000 lines of the Fuchsia Zircon kernel. It also says they rewrote 32,000 lines of SIMD code in a Rust port of the libgav1 video decoder, producing a version Google says runs 2.7x faster with identical output. Google adds that the large rewrites are still going through automated and manual audits before they reach production.

Why the pricing matters more than the benchmarks

At $2/$10, Argon's introductory price matches what OpenAI charges for GPT-6.1 Sol, which launched at DevDay the day before. Yahoo Finance made the same comparison. The announced post-introductory price of $4/$20 is the same list price Anthropic set for Claude Opus 5.5 last week. Both are well below the $10/$50 that GPT-6 Astra and Claude Fable 5.1 list at.

Put plainly, three labs have now put their frontier tier in the same $2 to $4 input band within about ten days. Model choice is becoming less about who is affordable and more about which model does your particular workload better per dollar. That is something you can only learn by testing on your own tasks.

Two caveats. The introductory price has no stated end date, so any cost model built on $2/$10 should also be run at $4/$20. And none of it applies to you yet: outside Fairwind, nobody can call the model today.

The 1M-token output budget changes the failure mode

The 1M output limit is the most practically distinctive item. Most production prompts are tuned around output ceilings in the tens of thousands of tokens. That is why long tasks get chunked, why agents are told to be terse, and why truncation is a common silent failure. Google's argument is that letting the model think and write "hundreds of thousands of tokens in a single trajectory" lets it solve harder problems in one pass.

The cost side is simple. One maximal response would cost $10 at the introductory output price and $20 after it, before any input or caching. A higher ceiling also means a runaway generation can get expensive before anyone notices. Once Argon reaches the API, explicit output caps, stop conditions and per-request budget checks will matter more than they did at 64K.

Safety posture and the reasoning-transparency ask

Google spends a large part of the post on safeguards. It says it monitors internal activations to detect misuse and runs red-team testing on the cyber and CBRN refusals. It claims leading results on Gray Swan's indirect prompt injection benchmark, again by its own account. It also describes monitoring Argon's chain-of-thought and actions to stop execution when the model goes beyond the user's intent.

One detail is easy to miss. Google says it took precautions not to feed misalignment-monitoring findings back into training, so that Argon's reasoning would not learn to evade the monitor. It also urges the rest of the industry to preserve reasoning transparency. That is a substantive position: it argues against optimizing chains of thought until they stop being informative.

What to watch

• A public model ID, API docs and rate limits, and whether the 1M output limit applies to every paid tier.
• The end date of the introductory pricing.
• Independent evaluations, especially on DeepSWE and the Vals Index, once access opens.
• Whether unguardrailed cyber access stays inside Fairwind, and how Google audits it.