Gemini 4 Argon is out. Here's what it changes for your workflows

Google released Gemini 4 Argon to developers and businesses worldwide on October 13, two weeks after announcing it as its most capable model yet. If your team runs automations on Gemini, or is picking a model for a new agent, you have about six weeks at the introductory price to find out whether it earns its place.

What happened

Google announced Argon on September 30 and gave it first to cybersecurity defenders in its Fairwind Program. It's built for long, multi-step work: real software engineering, legal and finance research, and finding and patching security flaws.

On October 13 it opened to developers and businesses through the Gemini API, Google AI Studio, the Antigravity coding tool and Gemini for business. Reuters sums up Google's pitch as "deep work" on complex, multi-step tasks.

Google also claims Argon scores significantly higher than OpenAI's GPT-6 Astra and Anthropic's Fable and Opus on several benchmarks, without publishing the head-to-head scores. Treat that as marketing until independent tests catch up.

The number to look at

Most launch charts measure coding or exam questions. One measures your kind of work: Zapier's AutomationBench, which tests models on realistic business automation tasks. Argon ranks first with 51.3%.

That's the best result any model has posted. It's also about half. On the other half, the model gets a step wrong, stops early or needs help. A better model raises the share of work you can hand off, and someone still has to check the output.

Two more numbers are worth knowing. Argon scores 77.9% on DeepSWE, a test of long software engineering tasks, and its output limit went up to 1 million tokens, from 64,000 before. Long reports, long migrations and long agent runs no longer stop halfway.

What it costs

At launch, Argon costs $2 per million input tokens and $10 per million output tokens. After the introductory period that becomes $4 and $20. According to Reuters, the higher prices start after November 30.

• Every processed token is billed, including the model's thinking tokens and the instructions behind tool calls. A step that thinks hard costs more than its visible answer suggests.
• Cached input is 95% cheaper, but only through implicit caching of repeated prompt prefixes of 6,144 tokens or more. Put fixed instructions and reference documents at the start of the prompt so they get cached.
• Explicit context caching isn't supported, so workflows built around it need rework.

For a workflow that runs thousands of times a day, the price after November matters more than the price today. Do your math with $4 and $20.

Switching takes more than a new model ID

Google's developer guide lists what changes when you move from Gemini 3.1 Pro to gemini-4-argon:

• One limit of 1,048,576 tokens now covers input and output together.
• A thinking_level setting (low, medium or high) replaces the old thinking budget. Medium is the default.
• temperature, top_p and top_k are ignored. If you used them to keep answers predictable, define a response schema instead.
• Prefilled model turns in chat histories have to go.

Argon runs in global, US or EU regions. If your contracts require customer data to stay in Europe, pick the EU endpoint before the first test, not after the first audit.

None of this is hard. All of it can break a workflow that gets switched over on a Friday afternoon without a test run.

What it means for your team

Use Argon where it's strong and keep it away from where it's only expensive. Most workflows mix two kinds of steps.

Short steps, like reading an order, sorting an email or pulling fields from an invoice, don't need a frontier model. Google's own guide recommends its lowest thinking level for classification, summarization and extraction, and a smaller model often does these just as well for less.

Long steps, like reconciling a month of transactions, checking a contract against a policy or fixing a broken integration, are where Argon's gains show up. Route those to Argon and leave the rest where they are.

Then test it on your own work. Take twenty real cases from last month, run them through your current model and through Argon, and compare accuracy, time and cost per case. Benchmarks show what's possible. Your own cases show what's worth paying for.

Once a step moves to Argon, watch it for the first few weeks. Log the input, the output and the cost of every run, and keep a person on the approval step for anything that reaches a customer or a ledger. You'll see where it fails before your customers do, and you'll know your real cost per case when the price changes.

Sources

• Google: Gemini 4 Argon, our next era of frontier intelligence
• Google Cloud: Developer's guide to Gemini 4 Argon
• Google DeepMind: Gemini 4 Argon model card
• Reuters via US News: Google launches flagship AI model Gemini 4 Argon for developers, businesses
• TechCrunch: Google releases Gemini 4 Argon, called its most powerful model yet