Gemini 4 Argon review: budget for the December price

What changed today

Google made Gemini 4 Argon generally available on October 13, 2026. Developers on the paid Gemini API tier can call it through AI Studio and the Gemini platform, businesses can reach it through Gemini for business, and AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search.

If you've been waiting to try it on real work, this is the week a lot of teams will start. That's exactly why the price deserves a minute before the excitement does.

Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. When the introductory period ends on November 30, 2026, Google's published rate becomes $4 and $20. Same model, same prompts, double the bill.

None of that makes Argon a bad pick. It just means the numbers you collect from an October pilot aren't the numbers you'll live with.

Who it's for

Google is pitching Argon at deep, multi-step work: large codebase refactors, performance tuning, document-heavy analysis in legal, finance and science, and security review. The launch post says Argon ranks first on the Vals Index, which scores models on finance, coding, legal and tax tasks. Its examples include a Chrome refactor across hundreds of files and an overnight optimization pass on Google's own Agent Development Kit.

Those are Google's examples, so read them as what the model is aimed at, not as a promise about your repo or your contracts.

The API gives Argon a 1 million token limit that covers input and output together, and Google points developers to a new continuation protocol for long runs. That matters if your job really is long: a big contract set, a sprawling codebase, a research pile. It's just overhead if your job is answering "where is my order."

So the fit, as published, is a team with a handful of hard, expensive problems. It isn't a team with a million small ones.

What it costs now, and what it costs in December

Here's the arithmetic on an invented but ordinary agent workload. Say each run reads about 40,000 tokens of context and writes about 8,000 tokens back, and you do 1,000 runs. That's 40 million input tokens and 8 million output tokens.

At the introductory rate, that's $80 for input and $80 for output, so $160. On December 1, the same 1,000 runs cost $160 for input and $160 for output, so $320.

Caching repeated context helps a lot right now, since cached input is 95% off. Google's launch post doesn't spell out the cached rate after November 30, so check the pricing page before you build the plan around it.

Two things push the number up further. Output costs five times as much as input, so a model that "works longer and harder" and writes more is spending at the expensive rate. And retries count: an agent that fails a step and tries again pays for its context twice.

The useful move is a boring one. When you write the business case, use $4 and $20. If it pays off at the list price, the next seven weeks are a bonus. If it only works at $2 and $10, you've found that out cheaply.

Where it falls short

Access is uneven. Ars Technica reported on October 13 that Argon in Antigravity and in AI Studio's coding tools is limited to AI Ultra subscribers, while Pro subscribers can generate code in the Gemini app's Canvas. Google is keeping Gemini 3.1 Pro available and hasn't said when, or whether, Argon reaches Antigravity on other tiers. If the plan was "the team will just use it in the IDE," check which plan each person is actually on.

It's the wrong tool for routine work. Classification, extraction, tagging and short summaries don't need a frontier reasoning model. Paying $20 per million output tokens for them is a habit worth breaking before it starts.

The safety numbers are Google's numbers. Google says Argon was trained to pause for human confirmation before irreversible steps, that shortcut attempts on its coding benchmarks fell to 2%, and that it leads the Gray Swan indirect prompt injection benchmark at 0.7% attack success. Those are encouraging, and the model card is worth reading. They don't replace your own permission boundaries, an approval step for anything that moves money or deletes data, and a record of what the agent did.

How it compares

Against Gemini 3.1 Pro, which Google is keeping around, the question is whether your task fails today. If 3.1 Pro already handles it, there's no reason to pay for the jump. If it stalls on long refactors or loses the thread halfway through a big document set, that's the gap Argon is aimed at.

Against the frontier models you may already run from Anthropic or OpenAI, or Meta's Muse Wonder, which landed a day earlier with its own benchmark claims, the charts won't settle it. This week will be full of them. The comparison that holds is yours: the same 20 or 30 real tasks, the same pass-or-fail bar, and the cost per passed task at the list price, not the launch price.

For the cheap work, compare Argon against a small, fast model rather than another flagship. Most stacks that work end up with both, and a simple rule for which jobs go where.

The verdict, and what to do this week

Verdict: Gemini 4 Argon is worth a trial if you have long, hard, multi-step problems and you're already on Google's API, or could be. It isn't worth moving everyday automation onto it, and it isn't worth building a plan around a price that expires on November 30.

To get real value from the discount, run the trial now on a fixed set of tasks, record the cost per passed task at both rates, and decide before December whether Argon earns the list price. Our note on testing a model against your own jobs before you buy covers the pass-or-fail bar, and tying agent spend to a job and an owner keeps the December invoice from being a surprise.

If you'd like help deciding which jobs deserve a frontier model and which don't, that's the kind of work we do in AI applications and agents.

Related work: Spin:Market.