A price list for a model you still cannot buy
Google has already priced Gemini 4 Argon. By its own account, it has not opened the door.
The claims live in Introducing Gemini 4 Argon: A New Frontier in AI Capabilities, on the company blog. The post does not show a publication date, so the figures below are Google’s, from that announcement, not a dated result an outside lab has checked.
The rate in the text is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95 percent off the input price. Google calls the rate “introductory” and hangs a footnote on the word. The body of the post does not say what the footnote carves out. Treat $2 and $10 as a starting quote, not a number you can take to finance.
Access is the sharper fact. Argon is rolling out to a set of trusted cyber defenders through the Fairwind Program. Google says it is in the U.S. government’s voluntary pre-release access process, and that developers, enterprises, and consumers come later, after early-tester feedback and another pass on guardrails. No agency is named. No date is given. “As soon as possible” is not a quarter on a roadmap.
If a vendor slide already treats this as a pilot you can run on your tenant, it is ahead of the announcement.
The blog post never says DeepMind. The Fairwind page is on a deepmind.google domain. That is the connection the launch pages actually supply.
The pitch is a long job, not a sharper reply
Google aims Argon at long-horizon software engineering, at enterprise knowledge work such as legal and finance, and at cybersecurity defense. For AI for business, the question is whether an AI agent can stay on a task that used to take a team days. A chatbot that got better at a short answer is a separate question.
The mechanism the post actually offers is output length. Argon is listed with a 1 million token output limit, up from a previous 64K. Google calls that industry-leading. No rival caps appear in the post, so the superlative is Google’s to defend. The post does not describe another mechanism (tools, a planner, a test loop) for how the model holds a long job. The business idea under the limit is simpler: one long trajectory, instead of a stack of short answers glued together by your staff.
That bet only matters if the trajectory survives contact with your codebase, your filings, and your audit trail.
Inside Google, the examples are specific and still internal
Google says thousands of its own people are already using Argon for specialized coding, deeper research, and writing. “Thousands” is the published figure. Missing are the task mix, the time saved, and the defect rate.
Google says quantum researchers saw Argon beat a published baseline by 40 percent on spacetime resources (the post’s own parenthetical is “qubits × gates”) “in a matter of minutes.” The baseline paper is not named.
On fleet profiling telemetry, Google says agents freed over 300 tebibytes of memory once the change rolled out, roughly a third of a pebibyte. Estimated total savings sit in a range from 500 tebibytes to 1 pebibyte. The first figure is memory freed so far. The second is an estimated range, not a second measurement that contradicts the first. Both are company estimates, not an audited result. Useful as a picture of what a long-running agent might do to infrastructure cost. Useless as a number to paste into your own business case.
The coding stories are the ones engineering leaders will forward, and the caveat is in the same breath. Google describes C and C++ to Rust migrations ranging from tens of thousands of lines in libraries such as re2 and libgav1 up to more than 800, 000 lines for the Fuchsia Zircon kernel. Those rewrites are still in automated and manual auditing, emulation testing, and review. They are not cleared for production.
On libgav1, agents replaced 32, 000 lines of hand-tuned low-level code. The memory-safe Rust decoder is said to run 2.7 times faster than the existing Rust port, with identical video output, and closer to optimized C++. Google does not claim it beat the C++ version. Faster than a slower port is a real result. It is not a victory over the code you already trust in production.
The numbers, and what each one measures
Every figure here is written in the text of Google’s post, not read off a chart, and none is paired with an outside lab’s confirmation.
On DeepSWE v1.1, a long software-engineering task, Google reports 77.9 percent and calls it a new state of the art for real-world long-horizon work. On LVBench, a long-video test, it reports 91.7 percent and calls that state of the art as well. On CWE-bench v1, a test of known software-weakness classes, it reports a tie for first at 68 percent, building on 3.8 Flash Cyber’s result on CWE-bench v0. The other model in the tie is not named. No v0 score is given.
Then the claims without a number. On the Vals Index (sectors weighted by contribution to U.S. GDP), Google says Argon is leading, and that it is similarly leading on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark. No numeric scores are in the text. The GDP weighting describes how an index is built. It is not a claim that this model moves U.S. output.
On Zapier’s AutomationBench, Google reports first place at 51.3 percent for end-to-end execution across core business functions. Rank and absolute score are different facts. A mid-50s result can be number one and still leave a large share of workflows unfinished, depending on how the benchmark scores a run. That is the honest frame for AI automation: leading is not the same sentence as handing it the close process on Monday.
Defensive cyber, on a shorter list
Google says Argon was trained to find vulnerabilities, validate them, and patch them, and that trusted defenders and internal Google teams will get the model “without cyber guardrails.” Do not translate that into “no safety controls.” The post does not say whether cyber-specific filters are lifted so defenders can inspect exploit-like material, or whether something broader is turned off. Until Google defines the phrase, it is an open question, not a spec.
That access rule is the product decision. A model positioned for defense is being issued by trust, to a circle Google selects, while a voluntary government pre-release process is still running.
Wiz (the cloud-security company Google has agreed to buy) is, Google says, using Argon in a program called Scan for Good. In an early demonstration, Google says Argon found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, a risk it says earlier frontier models missed. No vendor, product, or CVE is named. Because Wiz is not an outside lab in this story, the finding is not independent confirmation. An unnamed critical bug is a reason to pay attention, not a case study you can diligence.
An internal Google vulnerability benchmark is said to have found exposures across codebases in 20 languages. No score is published. Google also mentions a Wiz black-box web test. The account does not include a finished comparison.
What to do with a launch that is mostly a waiting room
If long-horizon AI agents are on your roadmap, steal the questions from Google’s own Rust story, not the headline. Who audits the diff? What does emulation cover? What still has to be reviewed by a human before production? A migration of more than 800, 000 lines that is not cleared to ship is a better planning template than a benchmark adjective.
Legal and finance buyers should ask for the Vals and Harvey numbers, not the word “leading.” Sales teams should not borrow the language either. The post has no AI-for-sales result. Security leaders should ask what “trusted defender” means, and what “without cyber guardrails” actually turns off, before a defensive model sits next to their data. Ask for the model card and the safety report before a marketing score becomes a risk acceptance. Those documents are where intended use and limits belong. The blog post is a launch note.
What Google published is a long-running agent, a token price, a narrow first audience, and a defensive capability issued by trust rather than by a checkout button. On Google’s telling, thousands of its own people are already using it. That is an official announcement with a limited rollout. Briefing the board as if Argon were a vendor you can sign this month, or ignoring it because the demo is not on your tenant yet, both misread what was actually published.
The questions worth asking before the roadmap moves
-
Can we deploy Gemini 4 Argon this quarter?
No. Google describes a phased rollout that starts with trusted cyber defenders in the Fairwind Program, while it is in a voluntary U.S. government pre-release process. Wider access has no public date.
-
What does it cost?
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, with cached inputs at 95 percent off the input price. The footnote on “introductory” is not explained in the post, so do not lock a budget to the rate.
-
What work is it actually aimed at?
Long-horizon software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense. The post’s stated way of staying on one long task is a 1 million token output limit, up from 64K. It describes no other mechanism.
-
Which scores are numbers, and which are just “leading”?
Google states DeepSWE v1.1 at 77.9 percent, Zapier’s AutomationBench at 51.3 percent, LVBench at 91.7 percent, and a first-place tie on CWE-bench v1 at 68 percent. Vals Index, Vals Finance Agent v2, and Harvey’s Legal Agent Benchmark are called leading, with no scores in the text. All of these are first-party.
-
Can we copy Google’s internal results?
Not as audited ROI. The 40 percent quantum gain cites an unnamed baseline, the memory figures are company estimates, and the large Rust migrations (including more than 800, 000 lines of the Fuchsia Zircon kernel) are still under audit and review.
-
What should security leaders take from the cyber rollout?
Google says trusted defenders and internal teams will get a model trained to find, validate, and patch vulnerabilities “without cyber guardrails.” The post does not define that phrase, so do not read it as “no safety controls.” The healthcare finding names no vendor, product, or CVE, and it comes through Wiz, which Google has agreed to buy.
-
So what is “Argon, ” exactly?
Google uses Argon as the name and does not say whether it is a size class, a codename, or a successor. The only earlier Gemini name in the post is 3.8 Flash Cyber, cited as a benchmark predecessor, not as a lineage.