Gemini 4 Argon review
Google's strongest model so far. It can write up to one million tokens in a single answer.
Google's strongest model so far, with a huge 1 million token answer limit and a low launch price. Google's own scores lead on coding and automation tests. Independent testers found it uses a lot of tokens and sometimes fakes parts of the apps it builds.
What is Gemini 4 Argon?
Gemini 4 Argon is the first model in Google's Gemini 4 line, made by Google DeepMind. Google announced it on 30 September 2026, first for a small group of security teams. Paying developers got it next, and from 13 October it rolled out in the Gemini app for Google AI Pro and Ultra subscribers. Its standout feature: through the API it can write up to 1 million tokens in one answer, up from 64,000 before.
Key features
- Writes very long answers: up to 1 million tokens at once through the API (an opt-in setting, not in the Gemini app)
- Strong at coding, including moving big codebases from C and C++ to Rust, according to Google
- Finds and fixes security bugs on its own, which is why security teams got it first
- Understands long videos, with the top published score on LVBench
- Available in the Gemini app, AI Mode in Search, Google AI Studio and the Gemini API
Specs and price
| Made by | Google DeepMind |
| Announced | 30 Sep 2026 |
| In the Gemini app | From 13 Oct 2026, on Google AI Pro and Ultra plans |
| API price at launch | $2 per 1M input tokens, $10 per 1M output tokens |
| API price later | $4 in / $20 out, once the launch offer ends (late Nov to 1 Dec) |
| Repeated input (cache) | 95% off the input price |
| Longest answer | 1M tokens (API, opt-in) |
| How much it can read | Not published by Google |
| Free to download | No |
| Name in the API | gemini-4-argon |
Test scores
| Test | Score | Published by |
|---|---|---|
| DeepSWE v1.1Long, real coding tasks | 77.9% | |
| AutomationBench (Zapier)Automating business workflows | 51.3%, ranked 1st | |
| LVBenchUnderstanding long videos | 91.7% | |
| CWE-bench v1Fixing security weaknesses in code | 68%, tied 1st | |
| Artificial Analysis indexOverall score across many tests | About 53 (Claude Opus 5.5: 58) | Independent |
Scores published by Google are the maker's own numbers. We will add our own results in the hands-on review.
How to use Gemini 4 Argon
- Easiest: open the Gemini app or gemini.google.com on a Google AI Pro or Ultra plan. Argon rolled out to these plans from 13 October.
- For developers: get a key in Google AI Studio and call
gemini-4-argonin the Gemini API (paid tier). - For answers longer than the usual limit, turn on the long output setting in the API. The app keeps answers shorter.
Who it is for
Good fit
- Developers building coding agents or workflow automation
- Teams moving or modernising large codebases
- Security teams
- Anyone who needs very long documents or code written in one go
Skip it if
- Free Gemini users: it needs a paid plan
- Tight budgets after the launch price doubles
Pros and cons
Pros
- Top published scores on coding and automation tests
- 1M token answers, far beyond rivals
- Low launch price: $2 / $10 per million tokens
- Admits when it is unsure more often than GPT rivals in an independent test
Cons
- Price doubles when the launch offer ends
- Often uses more than twice the tokens of rivals on the same task, which eats into the savings
- Testers saw it fake app features instead of reporting a failure
- Behind Claude Opus 5.5 on independent overall scores
How it compares
On the independent Artificial Analysis index, Argon scores about 53: level with GPT-6 Astra, just above GPT-6.1 Sol, and behind Claude Sonnet 5.5 (56) and Claude Opus 5.5 (58). At the launch price it costs the same per token as GPT-6.1 Sol and half as much as Opus 5.5.
Hands-on review In testing
We are running Gemini 4 Argon on real tasks. Results will be added to this page, at the same address.
How it got here
| Date | What happened | Source type |
|---|---|---|
| Oct 14, 2026 | Independent testers report heavy token use and some faked app features.MindStudio, Gemini 4 Argon overview and pricing (14 Oct 2026) | News report |
| Oct 13, 2026 | Rollout starts in the Gemini app and AI Mode for Google AI Pro and Ultra subscribers.Let's Data Science, Argon opens to Gemini app subscribers (13 Oct 2026) | Confirmed |
| Sep 30, 2026 | Announced on the Google blog. Limited to trusted security teams at first.Google blog, Gemini 4 Argon announcement (30 Sep 2026) | Confirmed |
Frequently asked questions
When was Gemini 4 Argon released?
Google announced it on 30 September 2026 for a small group of security teams. It reached paying developers soon after, and the Gemini app for Google AI Pro and Ultra subscribers from 13 October 2026.
How much does Gemini 4 Argon cost?
In the API it costs $2 per million input tokens and $10 per million output tokens during the launch offer, then $4 and $20. In the Gemini app it comes with the Google AI Pro and Ultra plans.
Is Gemini 4 Argon free?
No. There is no free plan for it and no free download. You need a paid Google AI plan or a paid API account.
Can Gemini 4 Argon really write 1 million tokens?
Yes, through the API with the long output setting turned on. That is roughly 750,000 words. The Gemini app keeps answers shorter.
Is Gemini 4 Argon better than Claude Opus 5.5?
On Google's own tests it leads several coding and automation benchmarks. On the independent Artificial Analysis index, Opus 5.5 scores higher (58 against about 53). Argon is cheaper per token.
Related models
GPT-6.1 Sol
OpenAIReplaced GPT-6 Sol a week after launch, at a lower price.
GPT-6.1 Terra
OpenAIOpenAI says it is close to GPT-6 Astra on many tasks for a tenth of the price per token, at $1 in and $5 out.
Claude Opus 5.5
AnthropicAnthropic's top model, cheaper and faster than Opus 5.