Google’s introductory API pricing for Gemini 4 Argon ends November 30, 2026. On December 1, the standard rate doubles from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. If you’re building anything on Argon — or even thinking about it — the next six weeks are the cheapest this model will ever be, and the work you do now decides whether December’s bill surprises you.
That’s the short version. The longer version matters more, because this isn’t just a price change. The same rollout split Google Search’s AI Mode into two tiers: paying subscribers get answers powered by Argon, while free users get answers from a different model. If you publish content, run a business, or build AI features, that split changes what your customers see — and what you should do before the deadline.
This piece is not another model-versus-model comparison. If you want the full flagship shootout, read our Gemini 4 Argon vs GPT-6, Astra, and Claude pricing comparison. This one is a deadline playbook: who should lock in usage before November 30, what to prototype now, how to budget for December, and what the AI Mode tier split means for publishers.
What Actually Changes on December 1
Let’s get the numbers on the table first, because everything else in this article hangs on them.
Google made Gemini 4 Argon generally available on October 13, 2026, through the paid Gemini API and Google AI Studio, plus the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers. The API launch price is introductory. According to Google’s API pricing documentation, as reported by TechRepublic, here is the full picture:
| Tier | Input / 1M tokens (through Nov 30) | Output / 1M tokens (through Nov 30) | Cached input / 1M | From Dec 1 (input / output / cached) |
|---|---|---|---|---|
| Standard | $2.00 | $10.00 | $0.10 | $4.00 / $20.00 / $0.20 |
| Batch and Flex | $1.00 | $5.00 | $0.05 | $2.00 / $10.00 / $0.10 |
| Priority | $3.60 | $18.00 | $0.18 | $7.20 / $36.00 / $0.36 |
A few details in that table deserve your attention:
- Output pricing includes thinking tokens. Argon is a reasoning-heavy model. The tokens it spends thinking through a problem before answering are billed at the output rate. That makes output, not input, the dominant cost in most real workloads.
- Cached input gets a 95% discount. When the model re-reads text it has already processed, you pay $0.10 per million tokens now, $0.20 after December 1. Caching is automatic for prompts of at least 4,096 tokens — you don’t set up your own caches, because explicit caching isn’t supported.
- Batch and Flex cost half the standard rate. If a job can wait, the Batch API charges half price. Priority costs more for faster, more reliable responses.
- There is no free tier. Argon is not available on the API’s free tier, and a Google AI Pro subscription does not by itself grant API access. You pay per token through the paid API, or you use it inside the subscription products described below.
One more number that shapes everything: Argon works with a unified 1-million-token budget that counts input and output together. The default synchronous output cap is lower (around 65,000 tokens), with up to roughly 1 million output tokens possible under a long-decode configuration. In practice, very long exchanges eat into the same budget on both sides, and long agentic runs can get expensive fast if you aren’t watching token usage.
Who Should Lock In Usage Before November 30
Not everyone needs to rush. The deadline matters most if you fall into one of these groups:
You should move now if…
You’re building an AI feature and haven’t picked a model yet. If your roadmap includes a coding assistant, a document-review tool, a research agent, or any multi-step workflow, the next six weeks are your cheapest window to test Argon against real workloads. A prototype that costs $300 in API calls now will cost $600 in December for the exact same prompts. That discount won’t come back.
You’re already running agentic workflows on another frontier model. Argon’s pitch is deep, multi-step work — complex coding, multi-file refactoring, finance, legal, and science tasks. If your current model struggles with long-horizon tasks or burns through retries, Argon is worth a head-to-head test while the test itself is half price.
You’re a freelancer or agency selling AI-powered deliverables. If you quote fixed prices for AI-assisted work — audits, code reviews, research reports, content systems — your margins depend on token costs. Prototyping now lets you lock in your process and your pricing before your cost base doubles.
You’re a publisher or SEO professional. The AI Mode tier split (more on that below) means paying subscribers already see Argon-powered answers. You need to know how your brand shows up in those answers before the model gets more expensive and more entrenched. That research costs you a $20-per-month Pro subscription, not API tokens — but the clock still matters, because the sooner you adapt, the longer you benefit.
You can probably wait if…
You’re a casual user who just wants to chat with a smart model. A Google AI Pro subscription at $20 per month already includes Argon in the Gemini app and AI Mode. The API price change doesn’t touch your subscription price. Nothing about November 30 forces you to act.
Your workload is simple Q&A or short-form generation. Argon is a heavyweight — don’t pay flagship prices for lightweight work.
You haven’t validated the problem yet. If you don’t have users, a prototype, or even a clear spec, a price deadline is not a reason to start spending. The deadline rewards people who already know what they’re building. It punishes people who buy tokens hoping inspiration follows.
What to Prototype Before November 30
Here’s my honest take: the worst way to use this window is to burn tokens on random experiments. The best way is to prototype the exact workflows you’d run in production, measure them, and walk into December with real numbers.
1. Your most expensive real workflow
Pick the single workflow that would cost you the most in production — the long document review, the multi-file refactor, the research synthesis. Run it on Argon now. Log three numbers: input tokens, output tokens (including thinking), and cached tokens. Those three numbers are your December budget. Without them, you’re guessing.
2. The agentic loop, not just the single prompt
Argon’s selling point is multi-step work: a task that takes ten, twenty, or fifty tool calls. A single clever prompt tells you almost nothing. Build the loop — the agent that plans, acts, checks its own work, and retries — and watch what it actually costs across a full run. Agentic loops are where thinking tokens pile up, and where the December doubling hurts most.
3. Caching, batch, routing, and your evaluation harness
Three architecture choices decide your December bill: structure prompts so stable instructions sit first and repeat verbatim (cached input costs 95% less — verify cache hits in your billing breakdown), build a Batch API version of every delay-tolerant job (batch in December costs what standard costs today), and prototype a router that sends easy work to cheaper models and reserves Argon for hard reasoning. Then build the unglamorous piece that makes all of it defensible: a small set of test cases with known-good answers, scored against Argon’s outputs.
The Math: What December Actually Costs
Let’s make this concrete with three realistic scenarios. All figures use standard-tier pricing and include thinking tokens in the output count.
Scenario 1: The solo developer. You run a coding assistant that processes about 5 million input tokens and 2 million output tokens per month. Through November 30, that’s $10 + $20 = $30 per month. From December 1, it’s $20 + $40 = $60. Manageable — but only if you measured it. If your agent actually burns 8 million output tokens because of long thinking chains, December costs $100, not $60.
Scenario 2: The small agency. You process 50 million input tokens and 30 million output tokens per month across client work. Now: $100 + $300 = $400. December: $200 + $600 = $800. If half your input is cache-eligible, you shave roughly $45 off the December input bill — small, but it adds up over a year. If you move offline jobs to Batch, December drops to roughly $400–$500. Same work, half the bill, just from architecture choices you make now.
Scenario 3: The heavy agentic pipeline. You run 200 million input tokens and 150 million output tokens per month, with long multi-step runs. Now: $400 + $1,500 = $1,900. December: $800 + $3,000 = $3,800. At this scale, the routing layer and batch migration aren’t nice-to-haves — they’re the difference between a sustainable product and a pricing page that doesn’t work anymore.
The pattern is the same at every scale: output tokens dominate, thinking tokens are the hidden multiplier, and the teams that measure now are the ones that don’t panic in December.
How to Keep Costs Down After December 1
The price doubles. Your bill doesn’t have to. Here’s the playbook, in order of impact:
Move everything delay-tolerant to Batch
Batch costs half of standard at every date. After December 1, batch rates ($2 input / $10 output) equal today’s standard rates. Any job that can wait — reports, bulk processing, backfills, evaluations — should move to Batch before December.
Design for cache hits
Keep stable prompt prefixes identical across calls. Reuse system prompts, instruction blocks, and reference documents verbatim. Every token that hits the cache costs one-twentieth of a fresh token. For agentic workflows that re-read the same context dozens of times, this is the highest-leverage optimization available.
Route ruthlessly, and cap thinking where you can
Reserve Argon for the work that actually needs deep reasoning; send everything else to cheaper models. And remember Argon exposes thinking controls (low, medium, high, with medium as the default). If your evaluation harness shows that low thinking passes your test cases for a given task class, use it — you pay for every thinking token, so reasoning depth is a budget setting, not just a quality setting.
Watch the unified budget, and re-price your own products
Because input and output share one 1-million-token budget, long conversations and long agent runs can run into the ceiling. Design workflows to summarize, truncate, or restart context before you hit it. And if you resell AI-powered work at fixed prices, update your pricing to reflect December’s costs — the worst outcome is holding November prices while paying December token costs.
The AI Mode Tier Split: What It Means for Publishers
Now the part that affects you even if you never touch the API.
Since October 13, Argon has powered AI Mode in Search — but only for Google AI Pro ($20 per month) and Ultra (starting at $99 per month) subscribers. Free users don’t get Argon in AI Mode. They get answers from a different model — Google has progressively restricted free-tier access, and free Gemini users are now limited to the lighter Flash-Lite model, as we covered in our piece on the free-tier Flash-Lite downgrade.
That creates something search has never really had before: a two-tier answer environment inside Google Search.
Here’s what that means in practice, and what to do about it.
Your brand may appear in one tier and not the other
Argon is built for deep, multi-step reasoning. It weighs evidence differently than a lighter model, and it may surface different sources. A brand that ranks well in free-tier AI answers might be invisible in Argon-powered answers, or vice versa. You can’t assume one result anymore.
What to do: Check both. If you have a Pro subscription, run your key brand and product queries through AI Mode as a subscriber. Then check the same queries in a signed-out or free-tier session. Note where you appear, where competitors appear, and what sources each tier cites. Do this weekly — the models and their answers shift.
Clear, verifiable facts win in the Argon tier
A reasoning-heavy model rewards pages where key facts are easy to find and confirm: prices, specs, dates, locations, credentials. If your important facts are buried in vague marketing copy, Argon has less to work with — and it may pick a competitor whose facts are stated plainly.
What to do: Audit your most important pages. Is your pricing stated plainly? Are your product specs in a clean table or list? Are your credentials, awards, and verifiable claims easy to find? This isn’t keyword stuffing — it’s making your facts machine-readable by a model that actually checks them.
Structured data matters more than ever
When two models generate answers from the same web, the one that reasons harder leans more on structured signals: schema markup, clean headings, FAQ sections, tables, and consistent entity names. If your site’s structured data is thin or broken, the Argon tier is where you’ll feel it first.
What to do: Make sure your key pages carry correct schema (Organization, Product, FAQPage, LocalBusiness where relevant). Keep your business name, address, and contact details identical everywhere. Fix the structured-data errors in Search Console before you worry about anything fancier.
You can’t measure the split with Google’s tools — yet
As of this writing, Google has not provided publishers with tools to measure the difference between the two answer sets. You can’t see an “Argon tier” report in Search Console. Instead, build your own lightweight tracking: a simple spreadsheet with your top 20 queries, checked weekly in both tiers, noting which sources each tier cites. Over time, you’ll see patterns — and patterns are what let you adapt.
The paying audience is the early audience
Pro and Ultra subscribers are, by definition, people who pay for AI — early adopters, professionals, and buyers. If Argon-powered answers favor your brand in that tier, you’re visible to a high-intent audience. If they don’t, you’re invisible to exactly the people most likely to act on a recommendation. That asymmetry is the real story of the tier split.
What Argon Is — and Isn’t
A deadline piece should be honest about the product, too. Here’s the grounded version:
Argon is Google’s first model in the Gemini 4 family, positioned for deep work: complex coding, multi-step agents, mathematics, science, finance, and legal tasks. Google cites internal gains including higher agentic code submission rates and fewer rollbacks, and says it topped the Vals Index benchmark. Treat all of those as Google-reported claims, not independent verification — the model has been broadly available for days, not months, and independent testing is still catching up.
Argon is not a general-purpose cheap model. It’s a flagship with flagship pricing, even at intro rates. It doesn’t generate images or audio, and it isn’t a drop-in replacement for every AI task. Teams that throw it at everything will pay flagship prices for commodity results.
Argon is not coming to the free tier on any announced schedule. Google has said nothing about free access. If your plan depends on Argon becoming free, you don’t have a plan — you have a hope. Build around the paid reality.
Argon is not the only way to get it. The Gemini app (Pro and Ultra), AI Mode in Search (Pro and Ultra), the paid API and AI Studio, Antigravity (Ultra only), and enterprise subscriptions all carry it in different forms. Pro subscribers get Argon in the app and in Canvas code generation, but Antigravity and AI Studio access are reserved for Ultra. Don’t pay $99 per month for Ultra if the $20 Pro tier covers your use case.
What NOT to Do Before November 30
Deadlines create bad decisions. Avoid these:
- Don’t stockpile tokens. There’s no token bank to fill. API pricing is pay-as-you-go; you can’t pre-buy November’s rate for December’s usage. Anyone selling you “discounted Argon credits” outside Google’s own billing is selling you something that doesn’t exist.
- Don’t sign annual contracts priced on intro rates. If a vendor quotes you AI-feature costs based on $2/$10 pricing, make sure the contract accounts for the December doubling — or you’ll absorb the difference.
- Don’t rebuild your whole stack in six weeks. Prototype, measure, and plan. A rushed migration to Argon that you haven’t evaluated is worse than staying on your current model until you’ve done the work properly.
- Don’t ignore the tier split because you don’t use the API. If you publish content, the split already affects you. Checking your brand in both AI Mode tiers costs you an hour and a $20 subscription, not an engineering sprint.
- Don’t assume the intro price will be extended. Google has done this before with other models — intro pricing that ends on schedule. Plan for December 1 as a real date, not a negotiation.
Your Week-by-Week Action Plan
Week of October 19: Identify your one most expensive real workflow. Run it on Argon. Record input, output, and cached token counts. Check your brand in both AI Mode tiers.
Week of October 26: Build the agentic version of that workflow, plus a batch-mode version for offline jobs. Test your caching structure and verify cache hits in billing.
Week of November 2: Build your routing layer and your evaluation harness. Score Argon against your current model on your own test cases, not on public benchmarks.
Week of November 9: Make the build-or-stay decision with real numbers. If you’re staying on Argon, migrate delay-tolerant jobs to Batch. If you’re switching or splitting, validate the cheaper model against your harness.
Week of November 16: Re-price your own products and client quotes for December costs. Update budgets, alerts, and spending caps in your API console.
Week of November 23: Final checks. Confirm your Batch jobs run, your router works, your cache hits, and your spending alerts are set. Do not start a new migration in Thanksgiving week — that’s how outages happen.
November 30: Last day of intro pricing. From December 1, every standard-tier token costs twice as much. If your architecture is ready, it’s just another Monday.
Frequently Asked Questions
When does Gemini 4 Argon pricing double?
The introductory API pricing ends November 30, 2026. From December 1, 2026, standard-tier rates double from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Cached input rises from $0.10 to $0.20 per million tokens, and Batch, Flex, and Priority tiers double as well.
Is Gemini 4 Argon free?
No. In the Gemini app and AI Mode in Search, Argon is included for Google AI Pro ($20 per month) and Ultra (starting at $99 per month) subscribers. On the API, it’s paid-tier only with no free tier. Free users don’t get Argon in any form under the current rollout.
What’s the difference between the AI Mode answers free users see and the ones paying subscribers see?
Paying subscribers’ AI Mode answers are powered by Gemini 4 Argon; free users’ answers come from a different, lighter model. That means the same question can produce different answers, different cited sources, and different brand and product placements depending on the searcher’s subscription tier. Publishers should check their visibility in both tiers.
How much does a typical project cost on Argon after December 1?
It depends almost entirely on output tokens, which include thinking tokens. A solo developer using 5 million input and 2 million output tokens per month would pay about $60. A small agency at 50 million input and 30 million output tokens would pay about $800 at standard rates — less with caching and Batch. Measure your own token mix before budgeting, because thinking-heavy agentic runs can push output far above a naive estimate.
Can I avoid the price increase by using Batch mode?
Partly. Batch costs half the standard rate, so after December 1, Batch rates ($2 input / $10 output per million tokens) equal today’s standard intro rates. Batch works for delay-tolerant jobs like nightly reports and bulk processing, but not for real-time features that need instant answers.
Should I switch to Argon before November 30, or wait?
If you have a real workflow to prototype, prototype now while testing is half price — then decide with measured numbers. If you don’t have a validated use case, don’t rush; the deadline rewards people who already know what they’re building. Either way, plan for December 1 pricing as the real cost, and never build a business case on intro rates that expire.