Models
Google announces Gemini 4 Argon, but almost no one can use it yet
Google DeepMind announced Gemini 4 Argon on September 30, a model the company says leads several tests. Only a group of cybersecurity specialists is using it, so almost no one can check that.
The essentials
4 confirmed facts · 1 according to the source · 3 open questions
- Google DeepMind announced Gemini 4 Argon on September 30. Only a group of trusted cybersecurity specialists has access, through the Fairwind program.135
- Google gives no date for the public. It says it will open the model to developers, businesses and consumers as soon as possible, after adjusting its safeguards.14
- According to Google, Gemini 4 Argon can write up to one million tokens (pieces of words) per response, compared with 64,000 for earlier models.17
- According to the sourceGoogle says Argon scores 77.9% on DeepSWE v1.1, a test of long programming jobs, ahead of its rivals.12
- Google has set a launch price of $2 per million input tokens and $10 per million output tokens.1
Why it matters
The jump to one million tokens points to a different kind of use: fewer questions and answers, more long jobs the model does on its own. That is the reading from Xataka, which estimates it equals about 750,000 words and that reading them would take about 50 hours7. If so, reviewing the work will mean checking the result, not reading all of it.
Closed access comes at a cost: almost all the figures come from Google and, as Xataka points out, hardly anyone can check them from the outside8. The outside measurement Xataka reports, from Artificial Analysis, puts it level with GPT-6 Astra and behind the Claude models6. Until it opens up, its results are claims from the company selling it.
The details
The announcement is signed by Koray Kavukcuoglu, Google's chief AI architect and an executive at Google DeepMind13. He explains that a model like this has to be opened up in stages, and that the company takes part in the US government's voluntary process for giving access to models before they are released1. According to MarkTechPost, Wiz is already using it and found a critical flaw in a healthcare program used by hospitals around the world4.
Google says thousands of its employees already work with Argon1. As an example, it points to its agents, programs that carry out a job from start to finish on their own: they are moving code from C and C++ to Rust, another programming language, including the core of the Fuchsia system, which runs to more than 800,000 lines12. Those changes are reviewed with automated tests and by hand before they are used1. It also says the model finds and fixes serious security flaws on its own15.
The scores come with fine print. According to MarkTechPost, which goes through the comparison Google published, Argon wins 12 of 18 tests and ties one, but falls behind GPT-6 Astra or Claude Opus 5.5 in at least three4. Xataka, citing Bloomberg, reports that some employees find it less polished on certain programming tasks. Google denies this, and Sundar Pichai maintains that its teams use it a lot and with good results6.
The launch price is not final either. MarkTechPost and Xataka report that it will later rise to $4 per million input tokens and $20 per million output tokens46. According to Xataka, the firm Artificial Analysis estimates that each task in its index costs $1.99 with Argon, 60% of what it costs with GPT-6 Astra, and that without the discount it will rise to $3.98, more than with its rival7.
What we don't know
- When the public will be able to use it and whether it will reach Europe at the same time: Google gives no date.
- How long the launch price will last before the increase reported by MarkTechPost and Xataka.
- Whether the model performs in real work as it does in Google's tests: there are hardly any independent checks.
Background
According to Xataka, this is not an isolated case: Anthropic keeps Claude Mythos Preview in the hands of a few organizations, and the most capable models now go first to whoever each company considers trustworthy8. The competition is also about price: on October 2 we reported that OpenAI promises, with GPT-6.1 Sol, almost the same as GPT-6 Astra for a fifth of the price.
How this story has changed
- The central facts do not change: we have added the use Wiz is already making of it according to MarkTechPost, the outside measurement from Artificial Analysis reported by Xataka, the precedent of other restricted models and what changes for the reader.
- We published the story.
Sources
- 1Gemini 4 Argon: our next era of frontier intelligence
- 2Google announces Gemini 4 Argon AI model, but you can't use it yet
- 3Google announces Gemini 4 and says it’s so capable that only ‘trusted cyber defenders’ can have it right now
- 4Google DeepMind Unveils Gemini 4 Argon with 1M Output Tokens for Coding, Knowledge Work and Cyber Defense
- 5Google releases Gemini 4 Argon, called its most powerful model yet
- 6Gemini 4 arrasa en los benchmarks y vuelve a meter a Google en la carrera de la IA
- 7Gemini 4 puede escribir un millón de tokens de una sentada: el producto ya no es el chat sino el trabajo largo y autónomo
- 8Lanzar un modelo de IA ya no significa lanzarlo. Ahora significa decidir quién puede usarlo
- ·Gemini 4 Argon 🧠, inside OpenAI’s chip ⚡, distillation takedown 🕵️
- ·[AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output
Was this story useful? Yes Not really
This story was written in Spanish by an artificial intelligence system from the sources listed above, and translated automatically. Before publication, a program checks that every figure and every name appears in the sources, and that the translation keeps the same facts, figures and sources. It has not been reviewed by a person. Spanish original · How we work (in Spanish)