Skip to content
Last checked: October 3, 2026. Argon is not open to the public yet, so this review is based on Google’s announcement and early reporting, not our own hands-on testing.
Gemini 4 Argon is Google’s new frontier AI model and the first of the Gemini 4 generation. Google announced it on September 30, 2026, nearly a year after Gemini 3, and is aiming it at long, complex work: software engineering, legal and financial research, and cybersecurity defense. Here is what it does, how it scores, what it will cost and when you can use it.
Gemini 4 Argon at a glance
| Detail | Gemini 4 Argon |
|---|---|
| Announced | September 30, 2026 |
| Made by | Google DeepMind |
| Built for | Long, multi-step coding, research, legal, finance and security work |
| Output limit | 1 million tokens (up from 64,000) |
| Price (introductory) | $2 per million input tokens, $10 per million output tokens |
| Price (after the intro period) | $4 input, $20 output per million tokens |
| Access today | Trusted cyber defenders in Google’s Fairwind Program, plus Google’s own teams |
| Next in line | Paid Gemini API customers and Google AI Ultra subscribers |
What’s new in Argon
A 1 million token output limit
Most models can write tens of thousands of tokens in one reply. Argon can write up to 1 million, up from 64,000 on earlier Gemini models. In practice this lets it think through a hard problem for much longer, or produce very large outputs such as a full code migration, in a single run.
Built for long, multi-step work
Google designed Argon to keep reasoning across long workflows instead of answering one question at a time. It is already in daily use inside Google. A few examples Google shared:
- Data centers: teams of Argon agents found memory savings across Google’s servers that freed more than 300 TiB of memory.
- Code migration: Argon agents are rewriting C and C++ code into Rust, including a video decoder that ended up 2.7x faster than the earlier Rust version.
- Quantum research: it improved on a published result for a quantum computing problem by 40% in minutes.
Cybersecurity defense
Google trained Argon to find, confirm and fix software vulnerabilities on its own. Security company Wiz used it to spot a critical flaw in hospital software that exposed personal data, one earlier models had missed. Trusted defenders get Argon without its usual cyber restrictions, which is why access starts with them.
Reading charts and long videos
Argon can analyze professional charts, pull details from long videos and act on a series of documents. On LVBench, a long-video test, it scored 91.7%, which Google says is the best result so far.
Benchmarks
Every score below comes from Google. No independent lab has been able to rerun them yet, so treat them as a strong first signal rather than final proof.
| Benchmark | What it tests | Argon result |
|---|---|---|
| DeepSWE v1.1 | Long, real-world software engineering | 77.9%, a new high score |
| Vals Index | Work across finance, coding, legal and tax | No. 1 overall |
| AutomationBench (Zapier) | Running business tasks end to end | 51.3%, ranked No. 1 |
| LVBench | Understanding long videos | 91.7%, a new high score |
| CWE-bench v1 | Fixing security vulnerabilities | 68%, tied for first |
| Harvey Legal Agent | Legal research and drafting | Leading score |
| Gray Swan prompt injection | Resisting hidden malicious instructions | Most resistant model tested |
Google’s own comparison shows Argon ahead of OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 and Opus 5.5 on most tests. Early coverage also points out that Argon trails GPT-6 Astra on some terminal-based coding tests.
Safety
Google says it is strengthening four safeguards before a wide release:
- Misuse: refusing help with cyberattacks and chemical, biological, radiological or nuclear weapons, while still allowing legitimate research.
- Prompt injection: resisting instructions hidden in web pages or documents that try to hijack the model.
- Misalignment: monitoring Argon’s reasoning and actions, and stopping it if it goes beyond what the user asked.
- Secure testing: sealing off the environments where it is trained and tested.
Google is also taking part in the US government’s voluntary program for testing models before release.
Price
Argon will launch at $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period, which Google has not dated, prices double to $4 and $20. That makes it more expensive than Gemini 3.8 Flash ($0.75 and $3.75 introductory) but close to Gemini 3.1 Pro ($2 and $12).
When can you use Gemini 4 Argon?
Not yet, unless you are part of Google’s Fairwind security program. Google says the next group will be paid Gemini API customers and Google AI Ultra subscribers, then developers, businesses and consumers more broadly. It has not given dates. Until then, Gemini 3.8 Flash and 3.1 Pro are the best Gemini models you can use.
Our verdict
On paper, Argon is Google’s strongest model and its clearest answer yet to OpenAI and Anthropic, especially for long coding jobs and professional research. The 1 million token output and the internal results at Google are the most convincing parts of the launch.
- Strengths: top scores on long coding and business tasks, a huge output limit, strong video and chart understanding, and serious security features.
- Weaknesses: almost no one can use it yet, every benchmark comes from Google itself and the price doubles after the introductory period.
If you are a developer or business already on Gemini, plan to test Argon once API access opens. Everyone else can wait for independent reviews and wider availability.
FAQs
What is Gemini 4 Argon?
It is Google’s newest and most capable AI model, released September 30, 2026. It is built for long, complex work in coding, research, legal, finance and cybersecurity.
Is Gemini 4 Argon available?
Only to trusted cyber defenders in Google’s Fairwind Program and Google’s own teams. Paid API customers and Google AI Ultra subscribers are next.
How much does Gemini 4 Argon cost?
$2 per million input tokens and $10 per million output tokens at launch, rising to $4 and $20 after the introductory period.
Is Argon better than GPT-6 Astra and Claude?
Google’s benchmarks show Argon ahead on most tests, but not all, and no independent testing exists yet. Wait for outside results before deciding.
What happened to Gemini 3.5 Pro?
Google announced it in May 2026 but never released it. It moved its focus to Flash updates and then to Gemini 4.
Sources: Google’s Gemini 4 Argon announcement · TechCrunch · Axios
Want AI that writes and ships code for you today? Compare the best AI coding agents on Klyvoo.