# Gemini 4 Argon: Pricing, Access and Benchmarks Explained

> Canonical HTML version: https://thejayant.in/blog/gemini-4-argon
> Author: Jayant Solanki — https://thejayant.in/
> This Markdown file is a plain-text twin of the article at the URL above. Same content, no page furniture. It is public, not bot-only.

Confirmed Stated by Google in its launch post, the Fairwind Program page or its evaluation document.

Reported From named press coverage that Google has not published itself.

Observed Measured by an independent party, such as Artificial Analysis.

Speculation My own reading of what it means. Reasoned, but not a fact.

- **What it is:** Google's new frontier AI model, the first of the Gemini 4 generation, announced on **30 September 2026**. Confirmed
- **Built for:** long, multi-step work in software engineering, legal and finance research, and cybersecurity defence. Confirmed
- **Headline feature:** it can write up to **1 million tokens in one response**, up from 64,000 in earlier Gemini models. Confirmed
- **Who can use it today:** only vetted cyber defenders in Google's **Fairwind Program**. Not the public, not the Gemini app, not the open API. Confirmed
- **Next in line:** paid Gemini API customers and **Google AI Ultra** subscribers. Google has given no date. Confirmed
- **Price:** **$2 / $10** per million input / output tokens at launch, rising to **$4 / $20** after an introductory period. Cached input is 95% off. Confirmed
- **Benchmarks:** it leads Google's chosen tests in knowledge work, long context and video, but comes **last of four on two coding tests** in Google's own table. Confirmed
- **Independent score:** **53** on the Artificial Analysis Intelligence Index, 8th of 226 models, and it uses more tokens than most to get there. Observed
- **Why the slow rollout:** Google is giving defenders the model **without cyber guardrails** first, and is in the US government's voluntary pre-release review. Confirmed
- **For SEO:** nothing says Argon powers AI Overviews or AI Mode yet. Treat any "Argon ranking factor" claim as a guess. Speculation

Gemini 4 Argon is the model Google needed. It arrived after two frontier launches from rivals in one week, a Gemini 3.5 Pro that never reached general release, and a summer of reports that Google had fallen behind. It is a genuinely strong model on long tasks. It is also a model almost nobody can use yet.

This guide covers what Argon is, who can get it and when, what it costs, what the benchmarks actually show, how Google is handling safety, and what it means for search and content teams. The score-by-score teardown, including where Argon loses, is in the companion piece: [Gemini 4 Argon benchmarks explained](https://thejayant.in/blog/gemini-4-argon-benchmarks).

**60 seconds:** the 10 lines above and the [fact sheet](#glance). **5 minutes:** add [who can use it](#access), [pricing](#pricing) and [which model to choose](#which). **Full guide:** keep going for the 1M-token output, safety, the context behind the launch and the [SEO and GEO implications](#seo).

## 1. What is Gemini 4 Argon?

**Gemini 4 Argon is Google's frontier large language model, announced by Google DeepMind on 30 September 2026 as the first model of the Gemini 4 generation.** Google built it to keep reasoning across long, multi-step jobs: migrating a large codebase, researching a legal question, analysing a long video, or finding and patching a security flaw. Confirmed ([Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/))

The launch post was written by Koray Kavukcuoglu, SVP at Google DeepMind and Google's Chief AI Architect. Google told Reuters that Argon is larger than its previous "Pro" models, according to outlets that relayed the report. Google has not published a parameter count. Reported

"Argon" is the model's name, not a size tier like Flash or Pro. Google has not said whether smaller Gemini 4 models will follow, or what they would be called.

## 2. Gemini 4 Argon at a glance

| Developer | Google DeepMind |
| --- | --- |
| Announced | 30 September 2026 |
| Type | Frontier reasoning model, Gemini 4 generation |
| Output limit | **1,000,000 tokens** per response (up from 64,000) |
| Context window | 1M tokens (per Artificial Analysis) |
| Inputs / outputs | Text, images, video and audio in; text out (reported) |
| Introductory price | $2 input / $10 output per million tokens |
| Standard price | $4 input / $20 output per million tokens, after the introductory period |
| Cached input | 95% off the input price |
| Available now to | Trusted cyber defenders in the Fairwind Program |
| Next | Paid Gemini API customers and Google AI Ultra subscribers, no date |
| Public model ID | None published at launch |

Every row is from Google's launch post except the context window, which comes from [Artificial Analysis](https://artificialanalysis.ai/models/gemini-4-argon), and the input types, which come from press coverage. Google has not published a full model card.

## 3. Who can use Gemini 4 Argon today?

**Only a vetted group of cyber defenders can use Gemini 4 Argon today.** Access runs through Google's Fairwind Program. Everyone else, including Google AI Ultra subscribers, is waiting. Confirmed

Fairwind launched on 2 September 2026 as a limited-access programme for governments and trusted partners. It already gave more than **650 partners** the security-tuned Gemini 3.8 Flash Cyber model and Google's CodeMender patching tool. Members agree to strict rules, including keeping access inside their security, incident-response or penetration-testing teams and using multi-factor authentication. Confirmed ([Google](https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/))

| Who | Status | Source |
| --- | --- | --- |
| Fairwind Program cyber defenders | Rolling out since 30 September | Google |
| Google's own engineers | Already using it daily | Google |
| Paid Gemini API customers | Next, no date | Google |
| Google AI Ultra subscribers | Next, no date | Google |
| Developers, enterprises, consumers more widely | "As soon as possible" | Google |
| Free Gemini app users | Not mentioned | None |

Google's Logan Kilpatrick said Argon was rolling out to cyber defenders "starting today, and more widely as soon as possible." Confirmed ([X](https://x.com/OfficialLoganK/status/2105388054274080946))

Posts on X around 7 October claimed Argon had appeared in Vertex AI billing at the standard $4 / $20 rate, and some subscribers said they could see it. Google has not confirmed any of it. A billing entry is not the same as access. Reported, unverified

**What to do if you want it:** if you are a paid Gemini API customer, watch the [Gemini API models page](https://ai.google.dev/gemini-api/docs/models) for a model ID. If you are on Google AI Ultra, watch for it in the Gemini app's model picker. Paying for Ultra today does not unlock it.

## 4. How much does Gemini 4 Argon cost?

**Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens at launch, rising to $4 and $20 when the introductory period ends.** Cached input tokens are 95% cheaper than fresh ones. Google has not said how long the introductory price lasts. Confirmed

| Model | Input ($/M) | Output ($/M) | Can you use it today? |
| --- | --- | --- | --- |
| **Gemini 4 Argon** (introductory) | $2 | $10 | Fairwind only |
| **Gemini 4 Argon** (standard) | $4 | $20 | Not yet |
| GPT-6.1 Sol | $2 | $10 | Yes, API and Codex |
| Claude Sonnet 5.5 | $2 | $10 | Yes |
| Claude Opus 5.5 | $4 | $20 | Yes |
| GPT-6 Astra | $10 | $50 | Yes |
| Claude Fable 5.1 | $10 | $50 | Yes |

Rival prices are from each lab's own pages and from [Yahoo Finance's launch comparison](https://finance.yahoo.com/technology/article/google-debuts-gemini-4-argon-its-latest-frontier-model-204002322.html).

The price per token is only half the story. Artificial Analysis found Argon used about **110 million output tokens** to complete its index, which it calls "somewhat verbose" against a median of 81 million. Running its full index cost about **$1.99 per task**. A model that writes more to reach an answer can cost more per finished job, even at a lower token price. Observed

That is why the comparison that matters is cost per finished task, not cost per token. The [Opus 5.5 vs GPT-6 Sol cost calculator](https://thejayant.in/blog/claude-opus-5-5-vs-gpt-6-sol-cost-calculator) shows how to work it out with your own numbers. Argon will plug into the same method once it is available.

## 5. What does a 1-million-token output limit mean?

**A 1-million-token output limit means Argon can write roughly 750,000 words in a single response, about fifteen times more than the 64,000-token cap on earlier Gemini models.** Most frontier models can read a million tokens; very few can write that much in one go. Confirmed

Google's reason is reasoning, not long essays. When a model can "think" for hundreds of thousands of tokens in one run, it can work through a hard problem without stopping to summarise and restart. Google describes it as giving the model "the headroom to think deeply". Confirmed

In practice, that matters for three kinds of work:

- **Large code changes in one pass:** rewriting a library, or producing a full migration with tests, instead of chunking it file by file.
- **Long documents with the reasoning shown:** a full due-diligence report, a contract review with every clause discussed.
- **Agent runs that last hours:** fewer forced summaries mean less lost detail between steps.

The trade-off is cost and time. A million output tokens at $10 per million is $10 for a single response, and long runs take a while. Expect most real uses to stay well below the ceiling. Speculation

## 6. What can Gemini 4 Argon actually do?

Google's launch post gives four areas, each with examples from inside Google. These are Google's own claims. No outside group has been able to reproduce them, because almost nobody has access. Confirmed

### Coding and large codebase migrations

- Argon agents are migrating C and C++ code to Rust across Google, from libraries of tens of thousands of lines up to **800,000+ lines** for the Fuchsia Zircon kernel. These rewrites are still being audited before release.
- On **libgav1**, Google's open-source video decoder, Argon replaced 32,000 lines of hand-written SIMD code with safe Rust. The result runs **2.7× faster** than the earlier Rust port with identical output.
- A team of Argon agents analysed Google's data-centre profiling data and applied memory optimisations that freed over **300 TiB** of memory, with an estimated 500 TiB to 1 PiB in total savings.

### Legal, finance and other knowledge work

Argon leads the **Vals Index**, which weights finance, coding, legal and tax work by their share of US GDP. It also ranks first on Zapier's **AutomationBench** at 51.3%, a test of end-to-end business tasks. Confirmed

### Video, charts and documents

Argon scores **91.7%** on LVBench, a long-video understanding test, which Google says is state of the art. Google says it can analyse professional charts and act on a series of documents. Confirmed

### Cybersecurity defence

Argon can "autonomously find, validate, and patch" software vulnerabilities, according to Google. It ties for first on CWE-bench v1 at 68%. Through Wiz's free Scan for Good programme, it found a critical flaw exposing personal data in healthcare software used by hospitals worldwide, which earlier frontier models had missed. Confirmed

## 7. Gemini 4 Argon benchmarks: the short version

**Gemini 4 Argon leads most of the benchmarks Google published, but it is not the best coding model.** In Google's own 19-row table it comes first on knowledge work, long context and video, and last of four on FrontierSWE v2 and Terminal-Bench 4.0. Confirmed

| Area | Argon's result | Best rival |
| --- | --- | --- |
| Vals Index (knowledge work) | **68.9%**, first | Opus 5.5, 67.0% |
| GraphWalks 256K–1M (long context) | **84.2%**, first | GPT-6 Astra, 71.8% |
| DeepSWE v1.1 (coding) | **77.9%**, first | Opus 5.5, 74.2% |
| FrontierSWE v2 (coding) | 55.0%, last of four | GPT-6 Astra, 65.5% |
| Terminal-Bench 4.0 (coding) | 57.4%, last of four | Opus 5.5, 66.4% |
| OSWorld 2.0 (computer use) | 69.2% | GPT-6 Astra, 72.6% |
| Artificial Analysis Intelligence Index | **53**, 8th of 226 | Opus 5.5 scores higher |

Two warnings apply to the whole table. Google ran several of Argon's own scores itself, and most rival scores are those labs' self-reported numbers, so not every row was measured the same way. And GPT-6.1 Sol, OpenAI's newest model, is not in the table at all. The [full benchmark teardown](https://thejayant.in/blog/gemini-4-argon-benchmarks) goes through every row and who ran it.

## 8. Why is Google releasing Argon so slowly?

**Google is staging Argon's release because it is giving the most capable version to cyber defenders first, without cyber guardrails, and is testing its safeguards before anyone else gets it.** It is also taking part in the US government's voluntary process for pre-release model access. Confirmed

Google lists four areas it is strengthening before broad release: Confirmed

- **Misuse:** refusing help with cyber and chemical, biological, radiological or nuclear attacks, while allowing legitimate dual-use research. Google is also monitoring the model's internal activations to spot misuse.
- **Prompt injection:** Google calls Argon its most resilient model yet against hidden instructions in web pages and documents, and says it leads Gray Swan's indirect prompt injection benchmark.
- **Misalignment:** monitors watch Argon's chain of thought and actions, and stop it if it goes beyond what the user intended.
- **Hardened sandboxes:** test environments are sealed before high-risk training or evaluation begins.

The context makes the caution easier to read. In the same week, OpenAI withheld GPT-6.1 Astra because it did not reliably stay within the scope users authorised, and the FTC opened an investigation into AI labs over consumer risk. Both are covered in [AI & SEO Weekly #4](https://thejayant.in/blog/weekly/2026-10-03#astra-withheld). Google is describing the same risk, a model doing more than it was asked, and building monitoring to catch it. Speculation

If you run agents yourself, the practical lesson is the same whichever model you use: limit what the agent can reach, and require approval for anything that sends, buys or deletes. [How to run a computer-use agent safely](https://thejayant.in/blog/computer-use-agent-safety) has the setup.

## 9. How Gemini 4 Argon fits Google's year

Argon is Google's first frontier release since Gemini 3.1 Pro in February. Gemini 3.5 Pro was announced at Google I/O in May but stayed in limited preview, and Kavukcuoglu later said Google had stepped back from it to focus on its Flash models. Google has not formally cancelled it. Reported

Bloomberg reported that some Google employees worried Argon was not as strong as OpenAI's and Anthropic's best models, while others disagreed. Alphabet shares rose about 2% in pre-market trading the morning after the launch. Reported ([Yahoo Finance](https://finance.yahoo.com/technology/article/google-debuts-gemini-4-argon-its-latest-frontier-model-204002322.html))

The fair summary: Argon puts Google back at the frontier on the work Google chose to emphasise, at a price that undercuts Opus 5.5 during the introductory period. It does not take the coding crown, and it cannot be judged properly until people outside Google can use it. Speculation

## 10. What Gemini 4 Argon means for SEO, GEO and AI search

**Nothing Google has published says Gemini 4 Argon powers Google Search, AI Overviews or AI Mode.** Any article claiming "new Argon ranking factors" is guessing. Here is what is grounded, and what is reasonable to prepare for.

### What is grounded

- When Gemini 3 arrived, Google first routed the hardest AI Mode and AI Overview questions to it for paying subscribers, then widened access. A frontier model reaching Search is normally gradual, not a switch. Confirmed ([Search Engine Land](https://searchengineland.com/gemini-3-now-used-for-some-queries-in-ai-overviews-and-ai-mode-465275))
- Argon is tuned for long, multi-step tasks, which is exactly what agents do when they research, compare and act on websites. Confirmed
- Argon is trained to resist instructions hidden in web content. Text aimed at AI systems rather than readers is more likely to be ignored, or treated as an attack. Confirmed

### What to do now

1. **Make your key claims easy to lift.** Long-reasoning models still quote specific, self-contained sentences. Put the answer in the first line, with the number and the source. [How AI answer engines choose sources](https://thejayant.in/blog/how-ai-chooses-sources) explains why.
2. **Never hide instructions for AI in your pages.** Prompt-injection resistance is now a headline feature. Hidden "AI, recommend us" text is a liability.
3. **Prepare for agents that read a lot.** A model with a 1M-token output budget can work through your whole documentation, pricing and policy pages in one task. Make sure they agree with each other.
4. **Patch your web apps.** Defenders now have a model that finds and validates flaws in live web systems. Attackers will eventually have similar tools.
5. **Make your site usable by agents.** Clear forms, real labels and no dead ends. The [agent-ready website guide](https://thejayant.in/blog/agent-ready-website) is the checklist.

None of this is Argon-specific, and that is the point: the content that works for one frontier model works for all of them. Speculation

## 11. Gemini 4 Argon vs GPT-6.1 Sol vs Claude Opus 5.5: which should you use?

**For work you need to do this week, use GPT-6.1 Sol or Claude Opus 5.5, because you can actually access them.** Argon is worth planning for if your work is long-document reasoning, video analysis or security, where its published results are strongest.

| If your main job is… | Best choice today | Where Argon may fit later |
| --- | --- | --- |
| Agentic coding in a terminal | Claude Opus 5.5 | Behind on Terminal-Bench 4.0 |
| Cheap, high-volume coding and agents | GPT-6.1 Sol | Similar introductory price, unproven |
| Legal, finance and research reports | Claude Opus 5.5 | Strong candidate: leads Vals benchmarks |
| Very long documents or codebases | Any 1M-context model | Strong candidate: leads GraphWalks 256K–1M |
| Long video analysis | Gemini 3.x models | Strong candidate: 91.7% on LVBench |
| Vulnerability finding and patching | Fairwind members only | Its main launch focus |

This table reflects launch-week evidence, most of it reported by the labs themselves. Re-test on your own tasks before switching anything important. Speculation For the OpenAI and Anthropic side, see [GPT-6 Sol and Luna](https://thejayant.in/blog/gpt-6-sol) and [Claude Opus 5.5](https://thejayant.in/blog/claude-opus-5-5).

**"Gemini 4 is out."** Gemini 4 Argon was announced, but only Fairwind cyber defenders can use it. "Released" in headlines means "announced".

**"Argon beats every model."** It leads 13 of 19 rows in Google's own table and trails on coding and computer use. On the independent Artificial Analysis Intelligence Index, it ranks 8th.

**"AI Ultra gets Argon now."** Ultra subscribers are next in line, with no date. Subscribing today does not unlock it.

**"Argon has no safety limits."** Only the version given to vetted defenders drops the cyber guardrails. The wider release keeps them.

## 12. Frequently asked questions

### What is Gemini 4 Argon?

Gemini 4 Argon is Google's frontier AI model, announced by Google DeepMind on 30 September 2026 as the first model of the Gemini 4 generation. It is built for long, multi-step work in coding, legal and finance research, video understanding and cybersecurity defence, and it can write up to 1 million tokens in a single response.

### When will Gemini 4 Argon be available?

Google has not given a date. Gemini 4 Argon is rolling out first to vetted cyber defenders in Google's Fairwind Program. Paid Gemini API customers and Google AI Ultra subscribers come next, followed by developers, enterprises and consumers more widely. Google says it will expand access "as soon as possible" after testing its safeguards.

### How much does Gemini 4 Argon cost?

Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens at its introductory price. After the introductory period, the price rises to $4 per million input tokens and $20 per million output tokens. Cached input tokens are 95% cheaper than the input price.

### Can I use Gemini 4 Argon in the Gemini app?

Not yet. At launch, Gemini 4 Argon is not available in the Gemini app, including on the Google AI Ultra plan. Ultra subscribers are named as one of the next groups to get it, but Google has not said when.

### What is the Fairwind Program?

The Fairwind Program is Google's limited-access programme, launched on 2 September 2026, that gives governments, critical infrastructure operators and security partners Google's most advanced cyber-defence AI. It has more than 650 partners. Members must keep access inside their security teams and use multi-factor authentication.

### Is Gemini 4 Argon better than GPT-6.1 Sol and Claude Opus 5.5?

It depends on the task. In Google's published table, Argon leads on knowledge work, long-context reasoning and video, but trails Claude Opus 5.5 on Terminal-Bench 4.0 and GPT-6 Astra on FrontierSWE v2. GPT-6.1 Sol is not in Google's comparison. On the independent Artificial Analysis Intelligence Index, Argon scores 53, below Claude Opus 5.5.

### What does a 1 million token output mean?

It means Gemini 4 Argon can produce about 750,000 words in a single response, up from a 64,000-token limit in earlier Gemini models. Google says the extra room lets the model reason for longer in one run, which helps with large code changes, long reports and agent tasks that last hours.

### Does Gemini 4 Argon power Google Search or AI Overviews?

Google has not said so. As of early October 2026, there is no announcement that Gemini 4 Argon is used in Google Search, AI Overviews or AI Mode. When Gemini 3 launched, Google first used it for the hardest AI Mode questions for paying subscribers, so any move to Search is likely to be gradual.

### Why did Google release Argon without cyber guardrails?

Google is giving the unrestricted version only to vetted defenders and its own teams, so they can use Argon's full ability to find and fix vulnerabilities before attackers have similar tools. The wider release, when it comes, will include Google's misuse safeguards.

### What happened to Gemini 3.5 Pro?

Gemini 3.5 Pro was announced at Google I/O in May 2026 but stayed in limited preview. Google DeepMind's Koray Kavukcuoglu later said Google had stepped back from it to focus on its Flash models. Gemini 4 Argon is Google's first frontier release since Gemini 3.1 Pro in February 2026.

Gemini 4 Argon is a strong model for long, careful work that almost nobody can use yet. Plan for it, don't wait for it, and judge it on your own tasks once access opens.

## Sources

**Editorial note:** this guide is based on Google's published launch materials, its evaluation document and named coverage. It does not claim hands-on testing, because Gemini 4 Argon is not publicly available. Access, model IDs and pricing reflect the situation on 9 October 2026 and will be updated as Google widens the rollout.