GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
Gemini 4 Argon Pricing: Intro Rate, Standard Rate, Real Cost
2026/10/01 ·

Gemini 4 Argon Pricing: Intro Rate, Standard Rate, Real Cost

Gemini 4 Argon pricing explained through input, thinking, cache hits and long decode. Learn how shared token limits shape a budget and find current reAPI rates.

Gemini 4 Argon costs $2.00 input and $10.00 output per million tokens at Google's introductory Standard rate through November 30, 2026; December 1 rates are $4.00/$20.00. Google's output category includes thinking tokens, and cached input is priced separately[1][4]. On October 23, 2026, the reAPI Gemini 4 Argon page displayed $0.80 input and $4.00 output per million tokens. That is a reAPI quote for this check, not a promise to follow Google's future schedule[5].

The cost question is more useful when it starts with a task. A short classification, a repeated agent turn over the same repository and a long research report use different amounts of input, reasoning and output. This guide explains those billing dimensions, the shared context limit and the continuation behavior that a budget needs to account for. The examples below are hypothetical token budgets, not measured workloads or price quotes.

TL;DR

  • Input, cached input and output are separate pricing dimensions. Google's output category includes thinking tokens[1].
  • Google's introductory pricing runs through November 30, 2026. Standard pricing starts December 1; current reAPI rates remain on the model page[1].
  • Input, thinking and output share a single 1,048,576-token context. You cannot combine a 100,000-token prompt with 1,000,000 generated tokens in that window[2].
  • One native non-streaming request returns up to 32,768 tokens; one stream returns up to 65,536. Longer generations use continuation requests[2].
  • Google documents implicit caching for prompts of at least 4,096 tokens. Explicit cache creation is unsupported for this model[1].

What counts toward a Gemini 4 Argon bill

The final answer is only part of token usage. Gemini 4 Argon can spend tokens reasoning before producing visible text, and Google includes those thinking tokens in the output pricing category[1]. A short answer therefore does not imply a small generation bill.

Billing dimensionWhat to inspect
Fresh inputPrompt and conversation material that did not receive a cache hit
Cached inputThe cached token count actually reported for that request
OutputGenerated answer and thinking usage, under the service's documented billing rules
Continued generationThe usage reported for every explicitly submitted continuation request

Keep the usage response with each request. Record the model ID, thinking setting, finish reason and whether the client continued the generation. That gives you a useful record when comparing runs with different output lengths. Counting characters in the visible answer cannot replace the reported token usage.

For an estimate where the service reports mutually exclusive token categories, multiply fresh input, cached input and billable output by their respective current rates, then divide each product by one million. Sum those terms using the service's billing and rounding rules. If input usage already includes cached tokens, subtract the cached subset before applying the fresh-input rate; otherwise the estimate counts those tokens twice. Follow the reAPI reference for the response and billing contract used here.

gemini api pricing

Google's October developer guide resolves a question left open by the September launch post: the introductory period lasts through November 30, 2026[1]. The old statement that there is no published end date is no longer correct.

For a Gemini API budget, compare the exact model and price period. The table below covers Gemini 4 Argon and Gemini 3.6 Flash, the two Gemini models offered by reAPI, using Google's Standard paid-tier rates in USD per 1 million tokens. Output includes thinking. Batch, Flex, Priority, tool charges, cache storage, and taxes are outside this table; Google's rate schedule is not a promise about reAPI's future tariff[4][6].

Google model and price periodInputOutput, including thinkingCached input
Gemini 4 Argon, through November 30, 2026$2.00$10.00$0.10
Gemini 4 Argon, starting December 1, 2026$4.00$20.00$0.20
Gemini 3.6 Flash, through December 31, 2026$0.75$3.75$0.075
Gemini 3.6 Flash, starting January 1, 2027$1.50$7.50$0.15

Google's Argon cache rate applies to implicit cache hits; explicit cache creation is unsupported. Flash has a separate cache-storage charge of $0.50 per million tokens per hour through December 31, 2026, then $1.00 starting January 1, 2027. Do not transfer Flash's storage model to Argon or count cached input again as fresh input[1][4].

At the October 23 check, reAPI's Argon Pricing section displayed $0.80 input and $4.00 output per million tokens, but no cached-input rate. The reAPI Gemini 3.6 Flash page still displayed “Pricing temporarily unavailable” and exposed no price offer. We therefore quote no current reAPI Flash price and do not reuse its July article price as a current discount. A missing cache rate is not a zero rate or permission to substitute Google's rate[5].

For 20,000 uncached input tokens and 5,000 billable output tokens, Argon's Google introductory Standard cost is (20,000 × $2 + 5,000 × $10) / 1,000,000 = $0.09. At the scheduled December rates it is $0.18. Using the observed reAPI input/output rates gives (20,000 × $0.80 + 5,000 × $4.00) / 1,000,000 = $0.036. These are token-only calculations with fixed counts, no cache hits, and no additional tool charges; they are not measured invoices or a guarantee of equal output quality[4][5].

On reAPI, Argon supports native Google GenerateContent requests. Flash uses reAPI's Chat Completions route with an 8,192-token output cap. Keep those request contracts separate when comparing budgets. Gemini 3.1 Pro is not offered by reAPI; the Flash vs Pro comparison explains that boundary, and the Flash usage guide covers its integration[6].

For a workload that will run across the transition, save two estimates using the rates applicable to each period. Keep token counts and task assumptions identical. Then change only the rates, so the result shows a pricing change rather than a mixture of pricing and workload changes. Do the same when comparing services: first verify that the request, thinking setting, cache assumptions and output budget mean the same thing on both.

Google's October 13 announcement separates access routes: developers use the Gemini API paid tier, while Google AI Ultra and Pro subscribers can use Gemini 4 Argon in the Gemini app and AI Mode in Search[3]. The developer guide's plan exclusions should not be generalized to consumer apps. Buying a consumer subscription does not prepay a reAPI request.

Why the shared token window changes cost estimates

Gemini 4 Argon has a unified 1,048,576-token context for input, thinking and output[2]. Input occupies part of that space before generation begins. The remaining budget must cover both the reasoning and the answer.

With exactly 100,000 input tokens, the arithmetic leaves at most 948,576 tokens for thinking and output combined. That is an upper bound within the context window, not a guarantee of a response that long. The per-request transport cap and your configured generation budget still apply.

Hypothetical taskInputCombined thinking and output budgetWithin the shared window?
Focused code review100,00020,000Yes; the model may finish earlier
Agent turn over reference material500,00050,000Yes; any cache hit must be confirmed from usage
Long report200,000600,000Yes overall; generation requires continuation beyond request caps
Oversized output plan100,0001,000,000No; the total exceeds 1,048,576

The table reserves a combined generation budget. It does not promise that all of that budget will become visible prose. On difficult tasks, thinking can consume a substantial part of it. If a report needs a particular length, test the task with enough space for reasoning and inspect why generation stopped[2].

Long decode uses more than one request

Google documents a cap of 32,768 generated tokens for one non-streaming request and 65,536 for one streaming request. Long decode can continue beyond those limits within the shared window[2]. Calling that “a million tokens in one response” hides a practical implementation detail that affects how you track usage.

Recent Google GenAI SDKs can continue automatically by default when the configured budget exceeds the per-request cap. Direct REST callers must manage continuation themselves[1]. reAPI does not automatically send continuation requests on the server. Your client decides whether to submit another request with the continuation information.

A client should record each request, decide whether the unfinished work is worth continuing, and retain the returned continuation token when present. If the model exhausts the configured generation budget, Google's guide instructs callers to raise that budget when resuming from the checkpoint[1]. Do not restart the full job blindly: restarting and resuming are different actions.

Google also notes that a difficult task can consume a whole segment thinking and return empty visible text with a continuation token[1]. An empty answer in that situation does not establish that no work happened. Inspect the finish reason and usage before deciding how the client should proceed.

Thinking and caching are different controls

The native thinking levels are low, medium and high, with medium as the default. Thinking cannot be fully switched off for Gemini 4 Argon[2]. Google recommends comparing levels on the same workload, then choosing the lowest setting that meets the required quality[1].

Lowering the output budget can truncate the answer as well as its reasoning. A more useful experiment keeps enough output space and varies the thinking level. Measure whether the answer meets your task's criteria, then compare token usage and elapsed time. A cheaper incomplete run is not a useful saving if your application has to start again.

Caching affects repeated input instead. Google documents implicit caching for prompts of at least 4,096 tokens and does not support explicit cache creation for this model[1]. Preserve a stable shared prefix when your workload permits it, but calculate savings from reported cache hits. A repeated prompt does not guarantee that every input token receives cached pricing.

This distinction also matters when comparing Gemini 4 Argon with another model. Do not assume that a separate cache-write charge, a cache lifetime or an explicit cache API from another product applies here. Use each product's documented mechanism and the usage fields returned by the service you actually call.

FAQ

How much does Gemini 4 Argon cost?

Checked October 23, 2026, Google Standard is $2.00 input, $10.00 output including thinking, and $0.10 cached input per million tokens through November 30; December 1 rates are $4.00, $20.00, and $0.20. The reAPI model page displayed $0.80 input and $4.00 output, with no cached-input rate listed. Use the applicable provider and period when estimating the bill[4][5].

Is Gemini 4 Argon free?

Google documents a paid API tier and separate Gemini app and AI Mode access for Ultra and Pro subscribers[3]. Those subscriptions are separate from reAPI's token billing; check the model page for current usage pricing.

When does introductory pricing end?

Google lists November 30, 2026 as the last day of introductory pricing, with standard pricing starting December 1[1]. That schedule is Google's; consult reAPI's live rate table for this service.

Are thinking tokens included in the cost?

Yes, Google's published output pricing includes thinking tokens[1]. Visible answer length alone is not a complete measure of generated usage.

Can I budget for one million output tokens plus a large prompt?

Input, thinking and output share 1,048,576 tokens. A 100,000-token prompt leaves at most 948,576 for thinking and output combined, before applying your generation budget and per-request caps[2].

Does reAPI continue a long generation automatically?

No. reAPI returns the response to your request; your client submits any continuation explicitly. See the request documentation before building a long decode loop.

Is Gemini 4 Argon cheaper than GPT-6 Astra or Claude Opus 5.5?

Compare current rates with the token usage of your own task. Different thinking budgets, cache hits and completion lengths can change the bill. The benchmark comparison covers the capability differences without treating price per token as price per finished task.

Budgeting for a finished task

Keep a small set of representative prompts and the criteria that make each result usable. Run the same evaluation when you change thinking settings, context size or models, and include every continuation in the record. That gives Gemini 4 Argon pricing a concrete meaning: the cost of work your application can actually accept.

References

  1. Google. What's new in Gemini 4 Argon. Retrieved October 14, 2026 from ai.google.dev/gemini-api/docs/generate-content/latest-model

  2. Google. Gemini thinking. Retrieved October 14, 2026 from ai.google.dev/gemini-api/docs/generate-content/thinking

  3. Google. Gemini 4 Argon: Built to solve your hardest problems. Published October 13, 2026; retrieved October 14, 2026 from blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon

  4. Google. Gemini API pricing — Argon introductory/standard rates and Flash promotion. Checked October 23, 2026. ai.google.dev/gemini-api/docs/pricing

  5. reAPI. Rendered model Pricing sections. Checked October 23, 2026. Gemini 4 Argon, Gemini 3.6 Flash. Flash pricing was temporarily unavailable at this check.

  6. reAPI. Current model integration and request contracts. Checked October 23, 2026. Gemini 4 Argon documentation, Gemini 3.6 Flash documentation.