Gemini 4 Argon has a context window of one million tokens, according to Google
Google states that Gemini 4 Argon has a context window of one million tokens. The Decoder also reports up to one million output tokens. Access remains restricted to selected cybersecurity defenders in Fairwind Program.
Google states that Gemini 4 Argon has a context window of one million tokens and supports long, multistep tasks. In addition to this input context, The Decoder also reports a separate limit of up to one million output tokens. According to this source, the model accepts text, images, video and audio, but generates text only. According to Google, it also has increased resistance to indirect prompt injection attacks and improved mechanisms for detecting misuse.
Google introduced the model on 30 September and is making it available to selected cybersecurity defenders through Fairwind Program. According to Google, the model was trained for defensive cybersecurity and can autonomously find, verify and fix critical software vulnerabilities. Public access is not yet available. Access is expected to expand after feedback has been evaluated; according to The Decoder, paying Gemini API customers and Google AI Ultra subscribers are expected to be among the first. No date has been announced.
The Decoder reports an introductory price of 2 USD per million input tokens and 10 USD per million output tokens. Later standard rates are expected to be 4 USD and 20 USD per million tokens, respectively; the source does not give a date for the transition. The discount on input tokens retrieved from the cache is expected to be 95 %, which corresponds to 0.10 USD per million tokens at the introductory rate and 0.20 USD at the standard rate.
According to Google, the model achieved 77.9 % on DeepSWE v1.1, helped save 300 TiB of memory in data centers, and its agents converted more than 800 000 lines of Zircon kernel code from C/C++ to Rust. According to The Decoder, the independent evaluation by Artificial Analysis gives the model 53 points in the Intelligence Index. In this testing, however, the model consumed an average of 62 000 output tokens per task, so a low per-token rate alone does not guarantee low costs per completed task.
Why it matters
Selected security organizations are gaining access to a model designed to find and fix vulnerabilities and handle long tasks involving code. It is not yet an available work tool for other users. In business evaluations, the cost per completed task will be important: long outputs can increase consumption, while reusing cached inputs carries a significantly lower rate.
Release card
Gemini 4 Argon
- Context
- The model supports up to one million input tokens and up to one million output tokens.
- Inputs
- The model accepts text, images, video and audio. It outputs text only.
- Price
- The announced introductory price is 2 USD per million input tokens and 10 USD per million output tokens. Later standard prices are expected to be 4 USD per million input tokens and 20 USD per million output tokens. Cached input tokens receive a 95 % discount on the applicable input price.
- Availability
- The model is being made available to selected security partners through Fairwind Program and is used by internal teams at Google. Broader access is expected to follow later, initially for paying API customers and Google AI Ultra subscribers; no date has been announced.
- DeepSWE v1.1 77,9 % According to Google, this result measures the ability to handle long-term software engineering tasks.
- AutomationBench 51,3 % According to Google, the score measures the execution of complete workflows across core business activities.
- LVBench 91,7 % According to Google, the result reflects the performance of the model in understanding long videos.
- CWE-bench v1 68 % According to Google, the result measures the ability of the model to fix security vulnerabilities.
- Artificial Analysis Intelligence Index 53 bodů At the High reasoning level, the model achieved the same score as GPT-6 Astra at the max setting and Claude Fable 5.1.
- According to Google, the model finds, verifies and fixes software vulnerabilities.
- The model helps with debugging and code migration.
- The model analyzes charts and the content of long videos.
- The model does not provide image, audio or video outputs.
In the Artificial Analysis Intelligence Index evaluation, the model scored 53 points, the same as GPT-6 Astra and Claude Fable 5.1. Claude Opus 5.5 scored 58 points and Claude Sonnet 5.5 scored 56 points.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
According to Google, the model has a context window of one million tokens.
-
The model was introduced on 30 September.; The model supports long, multistep tasks.; The model works with text, images and video.; Google reports increased resistance to indirect prompt injection attacks.; Google reports improvements to internal activation mechanisms for detecting misuse.
- The model was introduced on 30 September.
- The model supports long, multistep tasks.
- The model works with text, images and video.
- Google reports increased resistance to indirect prompt injection attacks.
- Google reports improvements to internal activation mechanisms for detecting misuse.
-
The model was trained specifically for defensive cybersecurity.; According to Google, the model autonomously finds, verifies and fixes critical software vulnerabilities.; According to Google, the model analyzes the content of long videos and charts.
- The model was trained specifically for defensive cybersecurity.
- According to Google, the model autonomously finds, verifies and fixes critical software vulnerabilities.
- According to Google, the model analyzes the content of long videos and charts.
-
The model supports up to one million output tokens.; The standard price is expected to be 4 USD per million input tokens later.; The standard price is expected to be 20 USD per million output tokens later.; Cached inputs receive a 95 % discount.
- The model supports up to one million output tokens.
- The standard price is expected to be 4 USD per million input tokens later.
- The standard price is expected to be 20 USD per million output tokens later.
- Cached inputs receive a 95 % discount.
-
Google participates in the voluntary US government process for access to models before their release.
-
Google reports a score of 77.9 % on the DeepSWE v1.1 benchmark.; According to Google, the model helped save 300 TiB of memory in data centers.; According to Google, agents converted more than 800 000 lines of Zircon kernel code from C/C++ to Rust.
- Google reports a score of 77.9 % on the DeepSWE v1.1 benchmark.
- According to Google, the model helped save 300 TiB of memory in data centers.
- According to Google, agents converted more than 800 000 lines of Zircon kernel code from C/C++ to Rust.
Two audiences, two different impacts
What this means
For individuals
Developers outside the selected participants in Fairwind Program cannot yet try Gemini 4 Argon or count on it as an available tool for everyday work. No date for broader access has been announced.
More practical updates →For a business
Selected security organizations can evaluate the model for the process of finding and fixing software vulnerabilities. When deciding whether to deploy it, they need to verify its success rate on their own tasks and the costs of completing them, because long outputs can outweigh the advantage of low per-token rates.
Risks and complianceCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.