Key facts
- Google launched its new frontier AI model, Gemini 4 Argon, on Wednesday.
- Argon is designed for cybersecurity defense, coding, research, and writing.
- The model scored 77.9% on the DeepSWE v1.1 software engineering benchmark.
- Argon achieved a 0.7% attack success rate on Gray Swan's prompt injection test.
- Argon supports an output limit of up to 1 million tokens, compared to 64,000 for previous models.
- The model is initially available to vetted security teams through Google's Fairwind Program.
Google parent company Alphabet launched its new frontier AI model, Gemini 4 Argon, on Wednesday, positioning it as its most capable model yet and a direct competitor to offerings from OpenAI and Anthropic. The company claims Argon demonstrates industry-leading performance across various benchmarks, particularly in cybersecurity, coding, and complex reasoning tasks.
On the DeepSWE v1.1 benchmark, which assesses AI's ability to handle long and complex software engineering jobs, Argon achieved a score of 77.9%. This performance surpasses rival models such as Claude Opus 5.5 (74.2%), GPT-6 Astra (74.1%), and Claude Fable 5.1 (67.4%). Additionally, Argon significantly expands the context window, supporting an output limit of up to 1 million tokens, a substantial increase from the 64,000 tokens of previous Gemini models.
Google highlighted Argon's cybersecurity capabilities, noting its 0.7% attack success rate on Gray Swan's Indirect Prompt Injection benchmark, an improvement over competitors. The model will initially be deployed to vetted security teams through Google's Fairwind Program, a cybersecurity initiative. Google stated that the model is being released without cyber guardrails to allow defenders to identify vulnerabilities more effectively, and that a wider release for paid API customers and Google AI Ultra subscribers will follow.
Google also noted that Argon leads on 12 out of 18 benchmarks, ties one, and trails on five, according to its own comparison table. The company is participating in the U.S. government's voluntary pre-release model access process.