Google has released Gemini 4 Argon, its newest AI model for coding and cybersecurity. Read about its capabilities, benchmark results, how to access it, and its release timeline since October 13, 2026.
Gemini 4 Argon is Google’s newest AI model for coding, professional work, and cyber defense. Google announced it on September 30, 2026, as a model for software engineering, enterprise knowledge work, and cyber defense. Argon is the first model in the Gemini 4 family. Developers and businesses have been able to use it since October 13, 2026.
The model follows Gemini 3.5, and Google chose to develop Gemini 4 instead of releasing Gemini 3.5 Pro. According to TechCrunch, Google focused this launch on coding and cybersecurity rather than on consumer chatbot updates.
Gemini 4 Argon’s Coding Capabilities
On the DeepSWE v1.1 benchmark, Argon scored 77.9 percent, ahead of GPT-6 Astra (74.1 percent) and Claude Opus 5.5 (74.2 percent). On FrontierSWE v2, however, Argon scored 55.0 percent, behind Astra (65.5 percent) and Opus 5.5 (62.3 percent). All of these figures come from Google DeepMind’s model card, so they remain vendor claims.
Google engineers use Argon for daily debugging, large codebase migrations, and algorithm design. In one test, an Argon agent worked autonomously overnight on google/adk-python. It cut memory session setup time by up to 40 percent in long sessions. Initial imports (cold imports) also became 21 percent faster.
Argon accepts text, images, video, audio, and PDFs as input and produces text as output. Its context window reaches 1,048,576 tokens for combined input and output. Standard output is capped at 65,536 tokens, but it can rise to 1 million tokens with the long decode feature. The API also supports code execution, function calling, and grounding with Google Search.
Gemini 4 Argon’s Cybersecurity Capabilities
Google trained Argon to automatically find, validate, and patch critical security vulnerabilities. On Google’s internal vulnerability benchmark, Argon found a range of flaws in codebases across 20 programming languages. Argon also scored 68 percent for vulnerability repair on CWE-bench v1. According to Jefferies, that score matches Grok 4.7.
Google released Argon first through the Fairwind Program, which serves governments and trusted partners. Trusted defenders and Google’s internal teams get a version without cyber safeguards. Projects approved by Fairwind use the model ID gemini-4-argon-cyber-permissive. The security firm Wiz is already using Argon through the Scan for Good program.
Google calls Argon its model most resistant to indirect prompt injection. NeuralTrust reports that Google has put in place a chain-of-thought and action monitor that can halt execution. DeepMind’s model card acknowledges that the team cannot rule out the possibility that Argon reaches the cyber risk threshold (CCL). As a result, Google prepared a safety case for Argon’s offensive cyber capabilities.
How to Access Gemini 4 Argon
Developers can use Gemini 4 Argon through the paid Gemini API and Google AI Studio with the ID gemini-4-argon. Google also offers it in Android Studio. Google AI Ultra subscribers can select Argon in Google Antigravity for complex planning and multi-agent orchestration. Businesses can access it through Gemini for business and Google Cloud.
Google AI Pro and Ultra subscribers can use Argon in the Gemini app. Antigravity is now also available directly in the Gemini app for macOS. According to The Rundown, the free Gemini API tier and the Google AI Pro plan do not include Argon API access.
Gemini 3.1 Pro users need to change their model ID to gemini-4-argon. The MINIMAL thinking level is not supported on Argon, and the Gen AI SDK must be updated to the latest version. Teams that want to test Gemini 4 Argon can start in Google AI Studio with a paid API account.
Find more useful information only at bacatulisan.com