Gemini 4 Argon:What the modelis actually built for

What is Gemini 4 Argon?

Gemini 4 Argon is the first model in Google's fourth Gemini generation, and it is explicitly built for tasks that run across many steps. The announcement of 30 September 2026 names three areas: software engineering on real codebases, enterprise knowledge work such as legal and finance, and cybersecurity defence. It was written by Koray Kavukcuoglu, who leads model development at Google DeepMind.

The one technical number that separates Argon from its predecessors sits on the output side. The model can now produce up to one million tokens in a single response, where earlier Gemini models stopped at 64,000. That reads like a footnote and is in fact the point of the model. Only with room for its intermediate steps can a model reason through a large problem in one pass, instead of chopping it into pieces that each lose context.

Google is fairly specific about its own usage. Argon agents are migrating C and C++ code to Rust inside Google; in the libgav1 video decoder they replaced 32,000 lines of hand-optimised code, and the result runs 2.7 times faster than the previous Rust port with identical video output. Another team freed over 300 terabytes of memory across Google's data centres. These are Google's own measurements on Google's own systems, and nobody outside has verified them.

What does Gemini 4 Argon pricing look like after the introductory period?

Google quotes two Gemini 4 Argon pricing tiers, and the second one sits in a footnote. At launch, a million input tokens cost two US dollars and a million output tokens cost ten. Once the introductory period expires, four and twenty apply, exactly double. Google has not published how long that period runs, and while that stays open we think the safer assumption is to budget at the higher rate from the start.

The doubling matters less, though, than where the money goes in the first place. Output tokens cost five times what input tokens cost, and Argon is the model designed to produce unusually long outputs. Put those two facts together and the most expensive property is precisely the one you would choose the model for. Take a month with ten million input and six million output tokens, roughly 200 longer agent runs: that is eighty US dollars at the introductory rate and one hundred and sixty after it. Three quarters of that sits on the output side.

Which is also where the obvious saving falls apart. Google discounts cached input tokens by 95 percent, so a cached million costs ten cents instead of two dollars. That is a real lever if you send the same long context repeatedly, a product manual or a contract archive for instance. It just applies to the smaller quarter of the bill. If half the input in our example comes from cache, the eighty dollars drops to roughly seventy.

And then there is the number pointing the other way. The evaluation platform Vals AI bills per run of its knowledge work index and reports 15.68 US dollars for Argon, already calculated at the post-introductory rate. The comparison models sit higher, between 21.34 and 32.14 US dollars (retrieved 8 October 2026). A high token price does not automatically produce a high invoice if the model needs fewer tokens to finish the same job. What that means for your workload only shows up in a comparison run on your own tasks.

Where the Gemini 4 Argon benchmarks promise more than they deliver

The most revealing line in the Gemini 4 Argon benchmarks is the last one. On Harvey's legal agent benchmark, which tests legal research and drafting, Argon scores 19.6 percent. That is the best result in the field by a wide margin, well ahead of GPT-6 Astra at 5.4. It also means the leading model fails four tasks out of five. Google names legal work as a target area, and in that very area the top score sits below twenty percent. Both things are true at once, and reading only the ranking takes away the wrong half.

The practical consequence is unglamorous. A model that performs that way on a domain benchmark replaces nobody in that domain. It produces a draft that a qualified person reviews, and that holds for contract drafting as much as for a tax opinion, regardless of how the numbers look in the press release.

The technical specifications deserve a second look too. Google states one million output tokens; the Vals AI profile for the same model lists 262,000 (retrieved 8 October 2026). Which figure applies to your account is something only your account will tell you. We would not base a plan on a number from an announcement without querying it once ourselves; with third-party limits, discrepancies tend to be the rule rather than the exception.

Finally, the most impressive examples in the announcement are Google's own. The 2.7 times faster video decoder, the 300 terabytes of reclaimed memory, the Fuchsia kernel migration: all measured by Google, on Google's code, with Google's engineers alongside. How the model behaves in a company with legacy software and no in-house engineering team is not in there.

What to settle before you request Gemini 4 Argon API access

Three things remain open around Gemini 4 Argon API access, and all three concern the setup rather than the model itself. This is the position as of 8 October 2026. First, region: Google has not published whether or when Argon will be offered in the EU, nor where processing happens. Second, contracts: for personal data you need a data processing agreement, and that attaches to the access path you use, not to the model name. Third, price: with no end date for the introductory period, the doubling is a deadline without a date, and those tend to land in months you were not watching.

A fourth item is not an open question but a lesson. Model names and prices are values somebody else sets and changes at will, and in most applications they appear in more than one place. Keep them in a single place and swapping a model takes minutes rather than days. That sounds like housekeeping right up to the first deprecation notice.

If you want to take the broader view: we have written up how to bring model spending down systematically under LLM cost optimization, and what AI agents can actually handle today under computer use agents.