Every AI lab wants a model that clocks in when you clock out. Google's newest pitch is basically "leave the hard engineering ticket with us overnight, and we'll have something for you with your morning coffee." Tempting. But as any manager of overnight contractors knows, the real question is what the invoice looks like.
Google's Flagship Learns to Pull All-Nighters
Google broadly released Gemini 4 Argon on October 13, calling it its "flagship reasoning model for complex engineering and scientific work." It's the first model in the Gemini 4 family and is built for long-running, autonomous multi-step agent tasks. Developers can reach it through the paid Gemini API, Ultra subscribers get it in Antigravity, and Pro and Ultra users get it in the Gemini app and AI Mode in Search.
Google's showcase examples are suitably ambitious: helping Chrome engineers complete an architectural refactor spanning hundreds of files in Chromium, speeding up Gemma 4 E2B inference 2.85x via vLLM, and cutting session preparation time by 40% in an overnight optimization of its Agent Development Kit. It ships with a shared input-output limit of about 1 million tokens, including thinking.
Cheap Tokens, With an Expiration Date
Pricing is $2 per million input tokens and $10 per million output tokens — but that's an introductory rate through November 30. On December 1 it doubles to $4 and $20. In an Artificial Analysis evaluation, Argon averaged $1.99 per task versus $3.26 for GPT-6 Astra, but it also used more than twice as many output tokens (roughly 62,000 vs. 27,000). Do the math: much of that cost edge rides on the discount, and the discount has a calendar reminder attached.
Benchmarks are a mixed bag, too. Argon posts 77.9% on DeepSWE v1.1, but trails GPT-6 Astra on FrontierSWE v2 (55.0% vs. 65.5%), and Claude Opus 5.5 scores higher on Terminal-bench 4.0. As The Neuron's analysis put it, the key measure is "whether the value of the finished task exceeds the combined cost of model use and review." That last word — review — is the line item most AI budgets forget. A model that writes 62,000 tokens of code overnight still needs a human to check it before it touches production.
Argon looks genuinely capable. Just pilot it like a new hire: real tasks, real review, and a close eye on the bill after the honeymoon pricing ends.
Weighing whether a long-running agent like Argon actually earns its keep in your business, or which model fits which job? That's the kind of AI integration math our team does every day — bring us your use case.
Source: Google Blog