Gemini 4 Argon Pricing: $2/$10 Now, Doubles Dec 1 (2026)

Google's frontier model is open to every paying developer, at half its list price until December 1

$2 in, $10 out, through November 30. That is what Gemini 4 Argon (gemini-4-argon) costs per million tokens since it became generally available in the paid Gemini API on October 13, 2026 - Gemini API changelog. On December 1 the price doubles to $4 input and $20 output, a date Google prints right next to the model ID in its migration guide - What's new in Gemini 4 Argon.

The launch drew 1,457 points and 1,038 comments on Hacker News by October 16 - Hacker News. The first wave of comments was about access, not quality, with readers reporting the model missing from their organization or region. Then one reader quoted Google's launch line and asked the question that matters for anyone building a product: what does it cost after the introduction, and from when? Another reader answered with Google's own pricing page: it doubles on December 1 - Hacker News comment. This dispatch covers what shipped, what a task really costs, the breaking changes to handle, and what to do before the price moves.

1. What Google Shipped on October 13

Google first showed Argon on September 30, but only to trusted cyber defenders in its Fairwind Program, promising wider access that would start with paid API customers and Google AI Ultra subscribers - Google. The October 13 release delivered that promise: Argon is in the paid Gemini API and AI Studio, in Antigravity for Ultra subscribers, and in the Gemini app for Ultra and Pro subscribers - Google India blog. What it does not have is a free tier, so every experiment is billed from the first token - Gemini API pricing.

The model shares one 1,048,576-token window across input, thinking and output, caps a synchronous answer at 65,536 tokens by default, and can write up to the full window with long decode - Gemini 4 Argon model page. It reads text, images, video, audio and PDFs, and it writes text. Every way of buying it doubles on the same date, so this is the table to plan against:

Two details in that table change real code. Output is billed including thinking tokens, and the cached rate applies only to implicit caching, because Argon drops the explicit cache that many Gemini apps were built on. Budget any feature you start today at the right-hand column: the introductory price lasts about seven weeks, and the feature will live much longer than that.

2. The Price That Matters Is Per Task, Not Per Token

At $2/$10, Argon matches the list price of Claude Sonnet 5.5, and from December 1 it matches the $4/$20 that Claude Opus 5.5 charges - Anthropic pricing. OpenAI's GPT-6.1 Sol also lists at $2/$10, while GPT-6 Astra sits at $10/$50 - OpenAI pricing. On the independent Artificial Analysis Intelligence Index, Argon scores 53, level with GPT-6 Astra and below Sonnet 5.5 at 56 and Opus 5.5 at 58 - VentureBeat.

A token price is only half of a bill. The other half is how many tokens a model spends to finish the job, and a reasoning model spends most of them thinking, which is billed as output. Artificial Analysis measured Argon at its high thinking level using an average of 62,000 output tokens per index task, against 27,000 for GPT-6 Astra - Artificial Analysis. That puts Argon at $1.99 per task today, 60% of Astra's $3.26, but $3.98 per task after December 1, about 1.2 times Astra. Even at the promotional price, it costs 2.7 times what GPT-6.1 Sol does per task.

Artificial Analysis broke that change down step by step, starting from what Google's previous Pro model cost per task:

Read it left to right. Argon's extra token use alone takes a task from $0.67 on Gemini 3.1 Pro Preview to $2.43, before any price change. The new token prices push it to $4.62, the deeper cache discount pulls it back to $3.98, and only the temporary 50% discount lands it at $1.99. The promotion hides the fact that Argon is a more expensive model to run than the one it replaces.

Apply that to your own workload with plain arithmetic. An agent task that reads 100,000 input tokens, half of them cache hits, and spends Argon's average 62,000 output tokens costs about $0.73 today and $1.45 from December 1. Output is more than 85% of that bill, so the levers that move your cost are the thinking level and the output cap, not the input price in the headline. We worked through the same logic for other models in our guide to setting the effort dial.

Where Argon earns its price is reliability. Its hallucination rate on AA-Omniscience is 15%, the lowest of any model scoring 45 or more on the index, against 51% for GPT-6 Astra, and it ranks first on AutomationBench-AA at 78%. It trails on Terminal-Bench 4.0, at 57% against 64% for Claude Sonnet 5.5 - Artificial Analysis launch analysis. A model that says "I don't know" more often is worth paying for in back-office work, and worth less in a command-line coding loop.

3. The Breaking Changes to Handle Before You Switch

Argon is not a drop-in swap for an app built on earlier Gemini models. Google's migration checklist lists changes that quietly alter behavior, and two of them can break a feature outright - What's new in Gemini 4 Argon. They all follow from one design choice: a model that thinks longer takes control of its own sampling, needs more output room, and runs longer per call.

That makes them easy to predict and easy to miss, because most of them fail without an error. A request that still sets a temperature returns a normal answer; a feature that capped output at 8,000 tokens returns a shorter answer, or none, once the model spends that budget thinking. Nothing in your logs says the model changed underneath you. Check each of these against your code before you change the model string:

• Sampling is ignored: temperature, top_p and top_k are dropped silently.
• Explicit caching is gone: cachedContents and caches.create are unsupported; only implicit caching, on prompts of at least 4,096 tokens, gets the 95% discount.
• Output caps need raising: Google recommends at least 32,768 max_output_tokens, because thinking counts against the cap.
• Calls run longer: Google advises HTTP timeouts of at least 15 minutes.
• Long answers need background mode: output past 64K tokens needs background set to true and a higher max_output_tokens.

The caching change is the one most likely to raise a bill without anyone noticing. An app that pinned a large system prompt in an explicit cache now depends on implicit cache hits, which only happen when requests share the same prefix, so prompt order matters: stable instructions first, user input last. Our prompt caching guide walks through that structure. The sampling change matters for any product that generates variants of the same thing, because asking for variety through temperature no longer does anything. The same trade, more thinking for fewer knobs, arrived with Anthropic's 5.5 models, as we covered in our Claude 5.5 migration guide.

The extra output room is the feature Google is proudest of. Its Google Cloud Tech channel published ten demos of 1M-token output on launch day, and it drew more than 260,000 views in three days:

Watch it with the cost math from section 2 in mind. Every one of those long outputs is billed at the output rate, so a 500,000-token answer costs $5 today and $10 from December 1, before any retries. Long output is a real capability, and it is also the fastest way to learn what Argon costs.

4. Why the Price Is Introductory

An introductory price is a lab paying for your migration. Moving a product to a new model costs engineering time, and Argon makes that cost concrete: remove the sampling settings, rebuild the caching, raise the caps and timeouts, and run the evaluations again. Seven weeks at half price is roughly the time a careful team needs for that work. Once the code is shaped around Argon, moving away costs the same work again, so the price can rise on a product that is already built around it.

That does not make the doubling certain. Anthropic launched Claude Sonnet 5 at an introductory $2/$10 through August 31 with a scheduled rise to $3/$15, then kept $2/$10 as the standard price - Anthropic pricing. Google, though, is running the same pattern on its fast tier: Gemini 3.8 Flash sits at an introductory $0.75/$3.75 until December 31 and doubles on January 1, 2027 - Gemini API pricing. Budget for the published date, and treat a reprieve as upside.

The window is long enough to answer the one question that decides whether Argon belongs in your product: what does a finished task cost on your own work, at each thinking level? Google tells developers to run their evaluations at low, medium and high and keep the lowest level that meets their quality bar - Gemini 4 Argon migration guide. Run that test now, while a mistake costs half. Nothing forces the move either: Gemini 3.1 Pro Preview has no shutdown date announced - Gemini API deprecations, so an app on it can measure Argon side by side instead of migrating under pressure.

5. The Bottom Line

A simple rule falls out of the numbers. If your task needs Argon's low hallucination rate or its automation scores, adopt it and budget at the December price, never the introductory one. If it does not, a model at the same $2/$10 that spends fewer tokens per task costs less today and much less after December 1. Either way, keep the model behind one setting so you can move when prices do, the pattern in our guide to model routing, and price your own product for the post-promotion rate, as we argued in pricing an AI product against token costs.

On Founden, Argon arrived in two steps. Founden's model catalog, which refreshes from vendor feeds, listed Argon at 22:59 UTC on October 13, and on October 15 its Google Pro line moved from Gemini 3.1 Pro Preview to Argon. Because Founden's model choices follow model families rather than pinned versions, a desktop build driven by Antigravity, the Google coder that joined Claude Code and Codex in Founden's desktop app on October 1, now defaults to Argon where the founder's own Google AI plan includes it. Google includes it in Antigravity for AI Ultra subscribers.

A price that doubles on a date only the lab chose is a decision every company built on that lab inherits. What a founder controls is whether moving to another model is a setting or a rewrite.

This dispatch reflects Gemini 4 Argon's pricing and documentation as of October 16, 2026. Google changes prices and limits often, so check the current pricing page before you commit a budget.