Gemini 4 Argon: Release Date, Price, Access & Benchmarks

Gemini 4 Argon was announced on September 30, 2026, but it is not a public Gemini model you can select today. Google is first giving approved cyber defenders access through its Fairwind Program. It says paid API customers and Google AI Ultra subscribers will be among the first groups in a wider rollout, but it has not given a date for that stage. Google has announced introductory API pricing of $2 per million input tokens and $10 per million output tokens; after the introductory period, it says those rates will be $4 and $20.

That distinction between announced and available is the most important part of this launch. Benchmark charts may suggest what Argon can do, but most developers cannot yet run their own workload through it. Here is what Google's announcement actually establishes, where independent testing adds useful context, and which questions remain open as of October 3, 2026.

Gemini 4 Argon at a glance

Question	Answer as of October 3, 2026
When was it announced?	September 30, 2026.
Can anyone use it?	No. Initial access is for approved Fairwind cyber-defense partners.
Who is next?	Google says paid API customers and Google AI Ultra subscribers will be early groups in its broader rollout. It has not announced dates.
What is the API price?	Google announced introductory rates of $2 per million input tokens and $10 per million output tokens, then $4 and $20 after the introductory period. It has not stated when that period ends.
What is the 1M-token claim?	Google describes an upper output token limit of one million for long trajectories. That is not a published one-million-token input context specification.

The first four rows come from Google's launch announcement; the access restriction is also explained on the Google DeepMind Fairwind page. The wording on output matters because many quick summaries call the one-million figure a “context window.” Google does not use it that way in its announcement.

When can I use Gemini 4 Argon?

There are two dates to keep separate. September 30 is the announcement date. Google has not given a general availability date for Gemini 4 Argon in the Gemini app, Google AI Studio or the Gemini API. Fairwind partners are the first users. The program prioritizes governments, critical-infrastructure operators and core technology platforms doing defensive security work; applicants are vetted and subject to access controls. This is not an ordinary waitlist for a personal Gemini account.

Google says it will expand access to developers, enterprises and consumers “as soon as possible,” starting with paid API customers and Google AI Ultra subscribers. That phrasing indicates an intended order, not a guaranteed launch week. Its public Gemini API model list did not include an Argon endpoint when checked on October 3. If a website offers an unofficial “Gemini 4 Argon API key” before Google lists access for your account, treat the claim cautiously and verify it against Google's own documentation.

What is new about the model?

Google presents Argon as a frontier model for long, multi-step work in software engineering, professional research, legal and financial tasks, and cybersecurity defense. It cites internal use for coding, research and large code migrations. Those stories illustrate where Google wants the model to help; they are not a guarantee that an outside team can reproduce the same result on a different codebase or workflow.

The standout technical claim is an output limit of up to one million tokens, up from a previous 64,000-token output limit described by Google. Output is what a model generates during a response or trajectory. Input context is what it can read at once. A higher output ceiling could help a long-running agent keep producing work without being stopped by a short generation cap, but it does not by itself tell us how much source material the public API will accept, how quickly it will run, or what a full-length task will cost. Those questions need the eventual model documentation and real use.

Google also says Argon can work across text, visual material and long video. It reports 91.7% on LVBench, a long-video understanding evaluation. For a buyer, the more useful question is whether the model can find the right evidence in your documents and present it with traceable citations. A general benchmark score cannot answer that on its own.

Gemini 4 Argon vs Gemini 3.8 Flash: which can you use now?

Google's current API catalog lists Gemini 3.8 Flash as a stable model with a documented endpoint, while it does not yet list an Argon endpoint. Flash is therefore a practical choice for a developer who needs to ship a Gemini-powered workflow today. Argon is Google's newly announced frontier model for harder, longer tasks, with limited early access. These are different product positions; a benchmark lead for Argon does not make Flash obsolete for routine, high-volume work where speed and total cost matter.

When Argon becomes accessible, compare both models on the same tasks and measure accepted results per dollar and per minute. In particular, do not compare a long-running Argon agent against one short Flash response and attribute every improvement to the model alone. Tools, repeated attempts and review time change the economics.

How good are the Gemini 4 Argon benchmarks?

In its launch post, Google reports 77.9% on DeepSWE v1.1 for long-horizon software engineering, 51.3% on AutomationBench for business workflows and 68% on CWE-bench v1 for vulnerability remediation, tied at the top of that security test. These are meaningful reported results, but the testing setup, tools, task selection and scoring rules matter. They do not prove Argon will be the best choice for every coding task or company.

An outside data point comes from Vals AI's published Argon evaluation. Vals reports a 68.90% result and first place on its GDP-weighted Vals Index at the time of its test, plus 65.40% and first place on Finance Agent v2. Its page also shows weaker placements: fifth on Terminal-Bench 4.0 and seventh of eight on CUA-bench. That mixed picture is more useful than declaring a universal winner. Vals lists a 262,144-token maximum output configuration for its evaluations; that test setting should not be mistaken for a refutation of Google's announced one-million-token output ceiling or for proof that every future API request will support it.

If you are comparing Argon with another model, use the same task set, tools, time budget, acceptance checks and human review process for both. Google scores and Vals scores answer different questions. A model that leads an index may still be slower, more expensive or less reliable on your particular work. Our agent evaluation guide gives a practical way to test outcomes before connecting customer data.

Gemini 4 Argon pricing: what would an API task cost?

Google announced introductory API rates of $2 per million input tokens and $10 per million output tokens. It says cached input tokens will receive a 95% discount from the input rate. After the unspecified introductory period, it says the standard rates will be $4 input and $20 output per million tokens. These are announced future API rates, not evidence that a public Argon endpoint is available today. Check Google's live API pricing page when access opens, because terms can change.

For an illustrative calculation, a task with 100,000 uncached input tokens and 10,000 output tokens would cost $0.20 + $0.10 = $0.30 at the announced introductory token rates. At the later announced rates it would be $0.40 + $0.20 = $0.60. This example excludes any other product, tool or infrastructure charges. A million-token output would carry a much larger bill if a workflow actually generated that much text; the maximum is not a sensible default target.

Why is access limited to cyber defenders first?

Google says Argon has strong defensive cybersecurity capabilities, including finding, validating and patching vulnerabilities. That same class of capability can be misused. Its Fairwind Program gives vetted defenders early access with restrictions on who can use the system and for what purpose. Google says it is also strengthening safeguards and conducting additional testing before broad availability. The company has described a staged safety rollout, not a public product launch with a hidden button.

For ordinary developers, the practical move is to prepare a small comparison set now: a few real tasks, clear pass/fail criteria, acceptable latency and a per-task spending limit. Once Argon is available in your account, run that same set against the model you already use. Avoid changing a production agent because a launch chart looks impressive.

Frequently asked questions

Is Gemini 4 Argon available in the Gemini app?

Google had not announced general app availability as of October 3, 2026. It named Google AI Ultra subscribers as an early group for a later rollout, without a date.

Can I use Gemini 4 Argon in Google AI Studio or through the API?

Not as a generally listed model on October 3. Google says paid API customers are among the first groups planned for wider access. Wait for an official model identifier and documentation before writing integration instructions.

Is Gemini 4 Argon free?

Google has announced paid API token prices and has not announced a free public Argon tier. Fairwind access is a separate, vetted program.

Does Argon have a one-million-token context window?

Google's launch post specifically calls out a one-million-token output limit. Do not substitute that number for an official input context specification. Vals lists a one-million-token context window in its model card, but public API parameters should be confirmed from Google when the model is listed.

Is it better than GPT-6 Astra or other frontier models?

It leads some published tests and does not lead all of them. Independent task results, access, speed and total cost matter more than one leaderboard rank.

Sources and reporting notes

Reported and checked October 3, 2026 using Google's September 30 announcement, the Google DeepMind Fairwind Program, the Gemini API model list, the Gemini API pricing page and Vals AI's evaluation. Google benchmarks are identified as Google's reported results; Vals figures are identified as Vals's own tests. The token-cost example is arithmetic based on Google's announced rates, not a bill from a public Argon request.