Gemini 4 Argon Pricing: What the Intro Rate Really Costs Before November 30
Google is charging half price for Gemini 4 Argon through November. Here is what a real task costs now, and what happens on December 1.
Jump to 8 sections
Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1. That is half price during the promotion and identical to Claude Opus 5.5's list price afterward. Budget at the standard rate, and measure tokens per task.
This comparison is for developers and teams choosing a frontier API for coding, analysis, or agents. Model Drop tracks launches and pricing, and our focus here is what a real workload costs, not only the rate card.
What Does Gemini 4 Argon Cost Right Now?
Google made Gemini 4 Argon globally available on October 13, 2026, according to Let's Data Science. It is offered through the paid Gemini API, Google AI Studio, enterprise services, the Gemini app, and selected Google AI plans. Google positions it for complex coding, knowledge work, cybersecurity, and scientific tasks.
The same report lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, running through November 30. Starting December 1, the standard price is $4 per million input tokens and $20 per million output tokens. An earlier pricing analysis, written before general availability, said Google had given no end date, so confirm the deadline on Google's own pricing page.
A token is a small chunk of text, roughly three-quarters of a word. Input tokens are what you send, and output tokens are what the model writes back. Output costs five times as much as input at both price levels, so verbose models get expensive fast.
Why would Google discount at launch? New models often use low introductory prices to win trial usage and developer attention. Teams that test during the promo can build habits and integrations that carry over after the price rises. That is a normal strategy, but it means the promo price is a poor guide to long-term cost.
How Does Argon Compare With Claude Opus 5.5 on Price?
On the rate card, Argon is cheaper only for a short window. Let's Data Science says Anthropic lists Opus 5.5 at $4 and $20 per million input and output tokens. Argon's intro rate is half that on both, and at the standard rate the two are equal, so the report calls the advantage "not a permanent advantage" under published rates.
| Model | Input per 1M tokens | Output per 1M tokens | Note |
|---|---|---|---|
| Gemini 4 Argon (intro) | $2 | $10 | Through November 30 |
| Gemini 4 Argon (standard) | $4 | $20 | From December 1 |
| Claude Opus 5.5 | $4 | $20 | Anthropic list price |
Our broader tracker of LLM API pricing in 2026 shows how these numbers compare with other models. Rates change often, so treat any table as a snapshot.
What Does a Real Task Cost?
A real task costs more than the rate card suggests when the model is wordy. Eesel's pricing analysis, written before general availability, uses Artificial Analysis data to estimate cost per task across its Intelligence Index. It lists GPT-6.1 Sol at $0.72, Claude Opus 5.5 at high effort at $1.82, Gemini 4 Argon at promo prices at $1.99, GPT-6 Astra at $3.26, and Argon at standard prices at $3.98.
The reason is verbosity. Eesel reports that Argon used about 62,000 output tokens per task, versus 27,000 for GPT-6 Astra. Output tokens are the expensive ones, so a model that writes more can cost more even at a lower rate.
Here is our own arithmetic from list prices. An agent run with 200,000 input tokens and 60,000 output tokens costs $1.00 at the intro rate and $2.00 at the standard rate. A long-document review with 800,000 input tokens and 10,000 output tokens costs $1.70 at the intro rate and $3.40 at the standard rate. These are illustrations, and they ignore caching and any long-context surcharges.
The same logic applies to your own usage patterns. A chat assistant that answers short questions will cost far less per request than an agent that reads many files and writes long plans. When you estimate a monthly bill, multiply the cost per task by your expected task volume, then add a margin for retries and longer-than-expected outputs. Retries are a hidden cost, since a failed attempt still bills every token it used.
Do Caching and Other Discounts Change the Math?
They can. Eesel lists cached input at $0.10 per million tokens at the promo price, which is 95% off, and notes that thinking tokens are billed as output. Eesel also says Google had not published batch, flex, priority, or long-context tiers for Argon at that time.
Caching helps most when you send the same long context repeatedly, such as a codebase or a policy document. If your workload has a large shared prefix, cached input can cut the input bill sharply. Our explainer on prompt caching shows how to structure prompts to benefit.
Output usually dominates the bill, though. In eesel's agent example, output made up about 97% of the cached cost. That means caching helps less for agent work than for document review, where input is large and output is small.
Is Argon Better or Just Cheaper?
It depends on the task. Let's Data Science says Google reports Argon leads on several knowledge-work, long-context, and coding evaluations, and that these are vendor-reported. An independent DataCamp comparison, it adds, found the models split: Argon led on more rows in Google's table, while Claude Opus 5.5 led on Terminal-Bench 4.0 and some terminal-heavy coding measures.
That split matches what we often see. A model can win on a broad average and still lose on the specific task you care about. Our overview of frontier model families in 2026 explains how to read these differences.
The safest approach is to test both on your own prompts. Compare quality, latency, output limits, tool reliability, and total tokens per finished task, as the report suggests. Public benchmarks cannot tell you how a model handles your repository or your documents.
Hallucination claims belong in the same bucket. Eesel notes that Argon's claimed hallucination rate is lower than some rivals, but that is a quality claim, not a pricing one. If lower error rates are true on your tasks, they can save money by reducing retries and review time. If they are not, the savings never appear. Only a test on your own data can settle it.
How Should You Plan Around the November 30 Deadline?
Plan on the standard price and treat the promo as a bonus. If your budget only works at $2 and $10, it will break on December 1. Doubling your estimate is a simple way to check whether the project still makes sense.
Use the promo period to gather data. Run your real workload through Argon for a few weeks and record tokens, latency, and quality. That gives you a true cost per task to compare with Claude Opus 5.5 or other models at standard rates.
Here is a short plan:
- Pick three to five real tasks that represent your workload.
- Run each on Argon and on your current model.
- Record input tokens, output tokens, latency, and pass rate.
- Multiply by the December 1 rates to see the true cost.
- Choose based on cost per successful task, not price per token.
What Else Should Buyers Watch?
Watch the details Google has not fully published. Eesel notes that batch, flex, long-context, grounding, and cache storage rates were not published for Argon at the time of its analysis. Check the current pricing page for each before you commit.
Watch availability, too. Early reporting described a restricted rollout, with access first for cyber defenders, then for paid API customers and Ultra subscribers. Let's Data Science says general availability arrived on October 13, so older posts about restricted access may be out of date.
Finally, watch your own usage. Model Drop's editors recommend setting spend alerts before you try any new frontier model, since a verbose model on an agent loop can burn through a budget quickly. For agent-focused options, see our comparison of AI coding agents.
One more tip: log everything during your test. Save the prompt, the model, the token counts, the latency, and a score for each answer. A simple spreadsheet is enough. When the promo ends and the price doubles, you will have the data to decide quickly whether to stay, switch, or split traffic between models by task type.
Conclusion
Gemini 4 Argon is cheaper than Claude Opus 5.5 only until November 30, and its verbosity can erase some of that edge. Quality splits by task, so neither model wins everywhere. The smartest move is to test on your own work and budget at the standard rate.
Your next step is to run a small side-by-side test this week, while the promo lasts. Model Drop will update this comparison if Google publishes new price tiers.
