Oct 14, 2026

Google Opens Gemini 4 Argon to Developers and Subscribers

Google's Gemini 4 Argon moves past its cyber-only preview, reaching API users and Pro/Ultra subscribers at $2/$10 per million tokens until Nov. 30.

Listen to this update

Google has made Gemini 4 Argon, its newest frontier AI model, broadly available. After two weeks in the hands of a small group of cybersecurity partners, the model is now open to developers, businesses and paying Gemini subscribers. Google describes it as built for long, multi-step "deep work" in coding, finance, law, cybersecurity and science, and is pricing it to win developers over from OpenAI and Anthropic.

From a cyber-only preview to general release

Argon first appeared on September 30. At that point Google limited it to vetted security firms in its Fairwind Program and pitched it mainly as a tool for defensive cyber work. TechCrunch reported at the time that Google said the model could find, validate and patch critical software vulnerabilities on its own.

The Tuesday, October 13 announcement widens access considerably. Koray Kavukcuoglu, Google DeepMind's chief AI architect, wrote the launch post. In it, Google says Argon is available globally through the paid tier of the Gemini API, AI Studio and Android Studio, and to enterprises through its business offerings. Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search. Access in Antigravity, Google's agentic coding environment, is limited to Ultra subscribers.

Reuters, as summarized by Slashdot, says Google is trying to make up ground after months of delays and a leadership shakeup at DeepMind. Google also calls Argon "just the beginning" of the Gemini 4 family, which means more models are on the way.

What Google says Argon can do

Argon comes with a context window of one million tokens through the Gemini API. Google says it improves on earlier Gemini models in spatial reasoning, mathematical precision and visual coding. It also says Argon follows style instructions more closely, hedges less and repeats itself less.

All of the performance figures below come from Google and have not been independently verified:

First place on the Vals Index, which scores models on finance, coding, legal and tax tasks
87% on the ERQA embodied-reasoning benchmark
More than 35% more agentic code submissions from Google's own engineering teams, and nearly a third fewer code rollbacks
A Chromium refactoring project that Google says would have taken months, finished in weeks
A 2.85x end-to-end inference speedup when Argon optimized a small Gemma model running on vLLM

Google also lists scientific results. These include faster code for a quantum-chemistry package, five new enzyme designs that showed catalytic activity in lab tests, and a robotics pipeline that taught Apptronik's Apollo humanoid to make seven cornhole shots in a row.

Pricing built to undercut

Until November 30, API pricing is $2 per million input tokens and $10 per million output tokens, and cached input is discounted 95%. After that date, prices double to $4 and $20. The introductory window gives teams about seven weeks to test Argon in real workloads at half the long-term price.

Google is also offering 10,000 scientists six months of free access to the highest usage tier of Antigravity. Eligibility and capacity limits apply. The Fairwind Program is expanding to hundreds more partners as well. Those defenders get a version of Argon without its usual cyber guardrails.

Safety claims

Google says it evaluated Argon under its Frontier Safety Framework and released it with a layered defense system for jailbreaks, monitoring and enforcement. The company reports these results:

External biology red-teamers unlocked fewer than 1% of the bio sub-tasks they tested.
Attack success on the Gray Swan indirect prompt-injection benchmark fell to 0.7%.
Attempts to take shortcuts on coding tasks fell to 2%.

Google also says Argon was trained to pause for human confirmation before irreversible actions. An automated monitor reads the agent's reasoning and actions in real time to block unintended workarounds. Google says it deliberately avoided penalizing the model's chain of thought during training so that the reasoning stays readable to the monitor.

Why it matters

When Argon debuted, TechCrunch noted that Google compared it favorably to OpenAI's GPT-6 Astra and to Anthropic's Fable and Opus models. Those comparisons are Google's own. Wider availability means developers can now test them on their own workloads. With a 1M-token context, aggressive introductory pricing and a focus on agent-driven coding, Google is going after the enterprise customers who decide which model they build on.

Also on Tuesday: Apple puts Siri in the living room

At a press event in New York, Apple announced HomeView, a $299 smart-home hub with a 7-inch display that can recognize up to six household members. It also announced a $299 Apple TV 4K with an A19 chip and a $149 HomePod mini 2. All three support the AI-powered Siri that Apple unveiled at WWDC. HomeView and Apple TV ship November 13, and HomePod mini 2 arrives October 30, according to Tom's Guide.