Last updated: September 2026
Key Takeaways
- Gemini 4 Argon is not free: trusted cyber defenders and early testers get it first, then paid API customers and Google AI Ultra subscribers, and Google has named no free tier or date for one.
- Paying does not buy the same privacy on every route. Gemini app chats, AI Ultra included, can be used to train Google's models unless you turn off Keep Activity and send no feedback, and human reviewers can see some chats either way; paid API prompts are not used to improve Google's products.
- Argon cannot run locally because Google has not released its weights. Google's open-weight option is still Gemma 4, which runs on hardware you own.
No. Gemini 4 Argon is not free, and as of September 30, 2026, most paying customers cannot use it either. Google is rolling it out first to trusted cyber defenders in its Fairwind Program, then to paid API customers and Google AI Ultra subscribers. Google has announced no free tier and no date for one.
Every figure below was checked on September 30, 2026, and this page updates as the rollout moves.
Who Can Use Gemini 4 Argon Right Now?
Only a small group can use Gemini 4 Argon today: trusted cyber defenders in Google's Fairwind Program, early testers, and Google's own teams. Google's announcement names paid API customers and Google AI Ultra subscribers as next and gives no date for that stage. On September 30, Argon did not appear on Google's Gemini API pricing page or its Gemini plans page.
| Who | Status on September 30, 2026 | What Google has said |
|---|---|---|
| Fairwind Program partners | Rolling out now | Governments, infrastructure operators, platforms, and security partners |
| Google's internal teams | In use | Internal projects cited in the announcement |
| Early testers | Testing now | Google is gathering their feedback while it adjusts guardrails |
| Paid Gemini API customers | Next, no date | Billed per token; not yet on the API pricing page |
| Google AI Ultra subscribers | Next, no date | Not yet on the Gemini plans page |
| Free Gemini, AI Plus, AI Pro, and AI Studio's free tier | No tier or date named | Wider access for developers, enterprises, and consumers as soon as possible |
Checked September 30, 2026 against Google's announcement, Fairwind post, API pricing page, and plans page.
Why Google is starting with defenders
Google ties the staged release to Argon's cybersecurity ability and says it is taking part in the U.S. government's voluntary process for pre-release model access. Fairwind, launched September 2 with more than 650 partners, limits use to each partner's internal security teams. Google has not said how long this stage will last.
How Much Does Gemini 4 Argon Cost?
On the API, Gemini 4 Argon will cost $2 per million input tokens and $10 per million output tokens at an introductory rate of unstated length, then $4 and $20. Cached input gets a 95% discount. Google also says Argon can write up to 1 million tokens in a single response, up from 64K on earlier Gemini models.
| Model | Input per 1M tokens | Output per 1M tokens | Max output per response |
|---|---|---|---|
| Gemini 4 Argon, introductory rate | $2 | $10 | 1,000,000 tokens (Google's figure) |
| Gemini 4 Argon, standard rate | $4 | $20 | 1,000,000 tokens (Google's figure) |
| GPT-6 Astra | $10 | $50 | 128,000 tokens |
| Claude Fable 5.1 | $10 | $50 | 128,000 tokens |
| Claude Opus 5.5 | $4 | $20 | 128,000 tokens |
List prices on September 30, 2026, before batch and cache discounts, from Google, OpenAI, and Anthropic.
At the introductory rate, Argon costs one-fifth as much per token as GPT-6 Astra and Claude Fable 5.1; at the standard rate it matches Claude Opus 5.5, per Anthropic's pricing page. Google's 1-million-token figure is nearly eight times the 128,000 tokens the others allow on standard requests.
What one maxed-out response costs
A single response that used the full 1-million-token limit would bill $10 in output at the introductory rate and $20 at the standard rate. Google's API pricing page counts thinking tokens as output for current Gemini models; if Argon follows that rule, its reasoning is billed too. Long jobs, such as the code migrations Google pitches, are where per-token bills grow fastest.
The consumer route: Google AI Ultra
For individuals, Google's named route is Google AI Ultra, its top consumer plan, which Google's plans page listed at $99.99 and $199.99 a month on September 30 without saying which tier includes Argon. Our breakdown of the Ultra tiers covers what each includes. The API is not a consumer path: Google's Gemini API terms reserve it for professional or business use.
Paying for Ultra Doesn't Buy API-Grade Privacy
Gemini app chats, including on a paid AI Ultra plan, fall under Google's Gemini Apps privacy terms, where keeping chats out of training takes a setting; the paid API keeps prompts out of product improvement automatically. Google's Gemini Apps Privacy Hub, updated September 24, says Google uses your activity to improve its services, including training generative AI models, and that human reviewers read some chats, which are then kept for up to three years even if you delete your activity. The page makes no distinction between free and paid plans.
The API terms run the other way: unpaid API prompts can be used to improve Google's products and read by reviewers, while paid API prompts are not used for improvement and are logged for a limited time for abuse checks and legal disclosures.
| Where you use it | Used to improve Google's AI? | How long it is kept |
|---|---|---|
| Gemini app, Keep Activity on | Yes, including model training; human reviewers read some chats | 18 months by default; reviewed chats up to 3 years |
| Gemini app, Keep Activity off or temporary chat | No, unless you send feedback; reviewers can still see chats for safety | 72 hours; any chat a reviewer sees, up to 3 years |
| Gemini API, unpaid tier | Yes; human reviewers may read prompts and responses | Not specified in the API terms |
| Gemini API, paid tier | No | Logged for a limited period for abuse checks and legal disclosures |
| Gemma 4 on your own hardware | No; nothing leaves your machine | As long as you choose to keep it |
From Google's Gemini Apps Privacy Hub (updated September 24, 2026) and Gemini API Additional Terms (updated April 28, 2026). Work and school accounts may be subject to different terms.
How to keep Argon chats out of training
If you reach Argon through Ultra, turn off Keep Activity in Gemini Apps Activity or use a temporary chat, and skip the feedback buttons on anything you want kept out of training. Google still keeps those chats for 72 hours, can have reviewers check them for safety, and keeps any reviewed chat for up to three years. Our guide to turning off AI features covers Google's AI Overviews and Gemini on Android.
Own the hardware your internet runs through
Keep Activity has an off switch. The remote access on a gateway your ISP rents you usually does not.
Rented gateways typically run TR-069 remote-management software, which lets the provider administer the router at the edge of your home network. Owning your router moves that control to you, and owning your modem ends the monthly rental fee. Our guide to owning your modem covers what changes.
The Argon You'll Pay For Isn't the One Google's Partners Get
Google says it will release Argon without cyber guardrails to trusted defenders and its own teams while it keeps adjusting guardrails with early testers, which implies everyone else gets a guarded version of the same model. Google has not published what those guardrails block.
Google is not alone. OpenAI says its GPT-6 Astra refuses more advanced cyber tasks, such as creating proof-of-concept exploits, while defenders can still use it for secure code review and patching, with looser safeguards planned through its Daybreak program. Anthropic routes some requests to other models, as our Opus 5.5 pricing breakdown explains.
For a buyer, the capability you pay for is set by the vendor's view of who you are, and the vendor can change it, sometimes silently, as the backlash over Claude Fable 5's silent safeguards showed. Only a model running on your own hardware is beyond a vendor's reach.
Is Gemini 4 Argon Worth Paying For?
The first independent scores put Argon at or near the top, but not alone there. Vals AI and Arena both rank it first, while Artificial Analysis scores it level with GPT-6 Astra and, per The Decoder, behind Claude Opus 5.5. All three published on launch day, before any public access.
| Evaluator | Measure | Gemini 4 Argon |
|---|---|---|
| Vals AI | Vals Index (finance, legal, tax, and coding tasks) | 68.9%, ranked first |
| Arena.ai | Text Arena (human preference votes) | 1,525 points, ranked first |
| Artificial Analysis | Intelligence Index | 53, level with GPT-6 Astra; Claude Opus 5.5 scores 58 |
Scores as published September 30, 2026; Claude Opus 5.5's index score as reported by The Decoder.
Artificial Analysis figures reported by Wccftech put Argon's cost at $1.99 per index task at the introductory prices, about 60% of GPT-6 Astra's, and roughly $3.98 once the discount ends. Argon also posted a 15% hallucination rate on the AA-Omniscience test, the lowest of any model scoring above 45 on the index.
Google's own chart is more favorable: across 18 benchmarks, Argon leads on 12, ties on 1, and trails on 5.
| Benchmark where Argon trails | Gemini 4 Argon | Leader | Leader's score |
|---|---|---|---|
| FrontierSWE v2 (agentic coding) | 55.0% | GPT-6 Astra | 65.5% |
| Terminal-bench 4.0 (agentic coding) | 57.4% | Claude Opus 5.5 | 66.4% |
| PostTrainBench (ML engineering) | 45.3% | Claude Opus 5.5 | 49.3% |
| Terminal-Bench Science 0.1 | 57.6% | GPT-6 Astra | 68.1% |
| OSWorld-2.0, offline subset, partial score (computer use) | 69.2% | GPT-6 Astra | 72.6% |
Source: Google's Gemini 4 Argon chart, September 30, 2026. Leader is the top score among the four models Google compared.
Google ties its long-horizon software engineering claim to DeepSWE v1.1, where Argon scores 77.9% to Claude Opus 5.5's 74.2%, yet two of the five benchmarks it trails are also agentic coding tests. Its widest leads are Harvey's legal agent benchmark (19.6% against 6.7%), GraphWalks at 256,000 to 1 million tokens (84.2% against 71.8%), and AutomationBench (51.3% against 42.5%).
The read: test Argon first for long-context and legal or finance agent work once you can reach it; for coding agents the evidence is split. Wait until it appears on Google's plans page before paying for Ultra to get it.
Can You Run Gemini 4 Argon Locally?
No. Google has not released Argon's weights and announced no open or on-device version, so there is no download and no self-hosted build.
Google's open-weight line is Gemma 4, released under Apache 2.0 in five sizes, per the Gemma 4 model card: E2B and E4B for phones and small machines, 12B, a 26B mixture-of-experts model with about 4 billion active parameters, and 31B. Our Gemma 4 launch coverage has the background, and our Gemini 3.6 Flash reality check maps each size to the hardware that runs it.
Gemma 4 is not a frontier model for long agentic work, but for drafting, summaries, document questions, and everyday coding it runs on your terms, and nothing you type leaves the machine. Our models-by-VRAM guide ranks open models at every memory tier, and our local AI hardware guide covers the machines.
The same logic applies to your modem. A rented modem is a monthly fee for hardware you never own, and buying one outright ends that fee. For cable internet, DOCSIS 3.1 is the newest standard sold at retail; check your ISP's approved list before you buy.
Frequently Asked Questions
Is Gemini 4 Argon free?
No. As of September 30, 2026, Gemini 4 Argon is available only to trusted cyber defenders in Google's Fairwind Program, early testers, and Google's internal teams. Paid API customers and Google AI Ultra subscribers come next, with no date given. Google has named no tier or date for free Gemini users, AI Plus or AI Pro subscribers, or AI Studio's free tier.
How much does Gemini 4 Argon cost?
On the Gemini API, Gemini 4 Argon will cost $2 per million input tokens and $10 per million output tokens at an introductory rate, then $4 and $20; Google has not said when the rate changes. Cached input is discounted 95%. For individuals, Google's named route is a Google AI Ultra subscription.
When will Gemini 4 Argon be available to everyone?
Google has not given a date. It says access will expand to developers, enterprises, and consumers as soon as possible, starting with paid API customers and Google AI Ultra subscribers. As of September 30, 2026, Argon did not appear on Google's API pricing page or its Gemini plans page.
Is Gemini 4 Argon the same as Gemini 4 Pro?
Probably, but Google has not said so. Pre-launch leaks described internal "Gemini 4 Pro" checkpoints codenamed argon, with a smaller output limit than the one Google announced. Google launched the model as Gemini 4 Argon and has not announced a separate Gemini 4 Pro, so treat pre-launch "Gemini 4 Pro" claims as unconfirmed.
What is Google's Fairwind Program?
Fairwind is a restricted Google cyber-defense program launched September 2, 2026, for governments, critical-infrastructure operators, technology platforms, and security partners. Partners get Google's cyber-capable models under strict rules, such as limiting use to internal security teams. It began with Gemini 3.8 Flash Cyber and more than 650 partners, and Argon is now rolling out through it.
Does Google train its models on Gemini 4 Argon chats?
In the Gemini app, including on paid AI Ultra, Google's privacy terms let it use your chats to improve its services, including training generative AI models, unless you turn off Keep Activity and send no feedback; human reviewers can see some chats either way. Through the paid Gemini API, Google says it does not use prompts or responses to improve its products.
Can I run Gemini 4 Argon on my own computer?
No. Google has not released Argon's weights, and there is no local or self-hosted version. Google's open-weight family is Gemma 4, licensed Apache 2.0 in sizes from E2B to 31B, which runs on phones, laptops, mini PCs, and consumer GPUs. It is smaller and less capable than Argon, but it runs entirely on hardware you own.
