Google’s Gemini 4 Argon has been available to the general public since Oct. 13, following a gated September release to cybersecurity defenders through the company’s Fairwind programme. Five weeks later, the model’s significance is clearer: it is Google’s attempt to make long-running, multi-step work the default use case for a frontier model, not an exotic edge case.

Argon is a new model rather than a fine-tune of Gemini 3, according to Google DeepMind’s model card. It accepts text, images, audio and video, with a combined input-output context window of up to one million tokens. Google positions it for software engineering, legal and financial analysis, cybersecurity defence and scientific research — tasks where the model has to hold a large amount of context and keep working after the first answer.

The pitch: fewer interruptions, more finished work

Google’s launch material emphasises behaviour as much as raw capability. Argon is said to answer more directly, hedge less, repeat itself less and avoid stock phrasing. On longer jobs it reports progress instead of going silent. That sounds cosmetic until you use these systems for real work: the most common failure mode in an agentic workflow is not a wrong answer but a model that loses the thread, stops early or asks a question it should have resolved itself.

The company also claims Argon outperforms competing frontier models across coding, cybersecurity, science, finance and law benchmarks, and says it ranked first on the Vals Index, which measures performance across economically significant sectors. Those results are company-reported and should be read alongside independent evaluations, particularly for tasks where benchmark contamination is hard to rule out.

Google has published an 87 per cent score on ERQA, a benchmark for embodied reasoning — understanding physical and spatial problems from images and descriptions. That is a robotics-adjacent result rather than a direct measure of enterprise performance, but it signals where Google expects the model to matter beyond software.

Pricing as strategy

Argon launched at an introductory $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95 per cent. That is dramatically cheaper than Anthropic’s new Fable 5.2 at $8 and $40, and it changes the economics of agentic workflows that make many calls over the same large context.

Cheap tokens do not automatically produce cheap outcomes. A model that needs three attempts, or that requires a human to review each step, can cost more than a pricier model that finishes. But the direction of travel is unmistakable: the frontier labs are competing on the cost of sustained reasoning, not just the cost of a single completion.

Distribution is Google’s advantage

Argon is available through the Gemini app and AI Mode in Search for Pro and Ultra subscribers, through AI Studio and the Gemini API, on Google Cloud, and inside Antigravity for developers. Google has also placed the Antigravity agentic coding interface in the Gemini macOS app, blurring the line between a chat product and a development environment.

For enterprises already buying Google Cloud, that bundling is the strongest argument. Procurement teams can attach a frontier model to existing cloud commitments, identity controls and data residency arrangements without a separate vendor review. The same logic applies to Microsoft’s Foundry and AWS Bedrock, which is why every major lab now ships through all three clouds.

The trust question

The harder problem is verification. Argon’s cybersecurity use case includes CodeMender, a tool for finding and patching vulnerabilities. That is valuable work, and it is also work where a plausible-sounding but incorrect patch can create a new vulnerability. Google’s staged rollout to vetted defenders before general availability reflects that risk, but the general release moves the judgement to customers.

Enterprises adopting Argon for long tasks will need to answer questions the launch materials cannot: Who reviews the model’s plan? What happens when it reports success incorrectly? Which actions require a human approval step? How is the context window kept free of data the model should not retain?

Google’s advantage in distribution and price is real. The constraint on adoption is no longer whether a model can hold a million tokens of context. It is whether an organisation can audit what the model did with them.

Reporting basis

This article draws on Google’s launch posts and Gemini 4 Argon model card, Google Cloud’s availability announcement, and reporting by The Verge, CNBC and the Google Developers Blog. Performance claims are attributed to Google unless otherwise stated.