Google DeepMind's Gemini 4 Argon costs $2 per 1 million input tokens and $10 per 1 million output tokens until 30 November 2026 — then it doubles to $4 and $20 from 1 December. If you are planning a build on this model, prototype, scope and test it before that deadline, because the same work will cost twice as much afterwards.

Announced on 30 September 2026 and generally available globally from 13 October 2026, Argon is positioned by Google for demanding work: coding, finance, legal, cybersecurity and science. It is available through the paid Gemini API and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. There is no free tier on the API, so every evaluation needs a budget.

Gemini 4 Argon: Build Now Before the Price Doubles on 30 November

What Gemini 4 Argon is — and what it is not

Argon is a frontier model from Google DeepMind, built for long, complex tasks: holding a large amount of material in mind, reasoning through it, and producing substantial, structured output. That makes it a candidate for production builds, not casual question-and-answer use.

Availability at a glance

• API: paid Gemini API, model ID gemini-4-argon; no free tier.
• App access: Google AI Pro and Ultra subscribers, in the Gemini app and AI Mode in Search.
• Inputs: text, images, video, audio and PDFs. Output: text.

Coverage of the release is summarised in ComputingForGeeks' Gemini 4 Argon release overview.

The 30 November deadline, in plain numbers

Until 30 November 2026, the API charges $2 per 1 million input tokens and $10 per 1 million output tokens. From 1 December 2026, those rates become $4 and $20 — so 1 million input plus 1 million output tokens costs $12 now and $24 afterwards.

That doubling hurts most for the workloads Argon is built for: long-context work with large inputs and long outputs. You need not launch by 30 November — but the cheap window for prototyping and evaluation closes then.

What a 1,048,576-token context window means in practice

Argon has a unified context window of 1,048,576 tokens, shared by input, thinking and output: thinking and output tokens draw from the same pool as input, so a long reasoning trace leaves less room for the answer. Default maximum output is 65,536 tokens, rising to around 1 million through long decode and continuation.

For a business build, that means giving the model the whole picture instead of a summary: an entire repository, a full set of contracts, a complete investigation file. It reduces fragile chunking — but not the need for discipline, since everything in the window still costs tokens. This is where AI agent development and AI automation change shape: agents that worked from summaries can now work from source material directly, if built around caching, capped reasoning and clear task budgets.

Where Argon earns its keep: five workloads Google positions it for

Google positions Argon for coding, finance, legal, cybersecurity and science. Its own cited figures — Google's claims, not independent verification — are DeepSWE 77.9%, LVBench 91.7%, CWE-bench 68% and AutomationBench 51.3%. Treat scores as a reason to test, not a promise about your workload: what does it do on your code, documents and cases?

• Coding: whole-repository refactors, migration and code review where the model sees callers, tests and configuration together. Scope it against a custom software development plan and a UK cost guide before assuming the model alone closes the project.
• Finance: long reports, filings and documentation reviewed as a whole, not in fragments.
• Legal: contract sets and case files where a clause in one document changes another's meaning.
• Cybersecurity: investigation material and vulnerability analysis spanning many files and logs.
• Science: long papers and literature where reasoning must stay grounded in the source text.

A sensible build plan before the deadline

You need not ship a finished product by 30 November. You need a defensible decision: what to build, what it costs to run, and whether Argon is the right model. A four-week plan:

1. Week 1 — pick one real workflow with a measurable outcome, such as triaging support documents.
2. Week 2 — prototype with real material, not demo data. Record input, output and cached-input tokens and latency for every run.
3. Week 3 — test the edges: the longest inputs, messiest documents and highest reasoning load.
4. Week 4 — scope and decide. Write down the per-task cost at both prices, the caching design and the thinking-level policy. Then commit, redesign or walk away — with numbers, not impressions.

That discipline applies whether the end product is SaaS development or a restaurant ERP feature — and if you are scaling an existing product, read how to scale a SaaS MVP without rebuilding first.

Four ways to keep the bill down

Four controls keep costs predictable:

1. Cache deliberately

Structure prompts so stable reference material is cached and only the new question changes per call. With 95% off cached input, this is the largest lever for document-heavy work.

2. Cap the thinking level

Argon offers three thinking levels: Low, Medium (default) and High. Higher thinking helps hard tasks but consumes context and budget. Default to Medium, reserve High for tasks that need it, and use Low for routine classification and extraction where accuracy holds up.

3. Budget per task, not per month

Set a token budget per task type — a maximum allowance per document reviewed or ticket triaged — and log actual usage against it. A monthly budget hides which workflow eats the money; a per-task budget shows it immediately.

4. Shorten outputs on purpose

Output tokens are the expensive side: $10 per million now, $20 after. Ask for structured, concise outputs — a verdict plus evidence, not an essay. These controls pair with good AI chatbot versus AI agent design and retrieval-augmented generation for business chatbots: use retrieval and caching to shrink what the model processes, and the big window for reasoning.

Test it on your work before you commit

Benchmarks describe a model; your evaluation describes your product. Before committing at post-December prices, run a small, honest evaluation: build a test set of 20–50 real examples with human-written expected outputs; run each thinking level (Low, Medium, High) and record accuracy, cost and latency; check failure modes, not just averages.

Further background is in Lindy's Gemini 4 Argon overview, but let your own test set decide. If it supports Argon, build with cost controls baked in from day one — whether that is ecommerce development, CRM development or a voice workflow from our guide on how to build an AI voice agent. Raise cost questions early: check AI calling agent costs, WhatsApp chatbot costs and mobile app development costs in the UAE before launch.

Start with one workflow, one test set and one costed decision — this week.

Frequently asked questions

When does Gemini 4 Argon get more expensive?

Introductory pricing of $2 per 1 million input tokens and $10 per 1 million output tokens runs until 30 November 2026. From 1 December 2026, the rates become $4 and $20 respectively.

How do I access Gemini 4 Argon?

Through the paid Gemini API (model ID gemini-4-argon), or as a Google AI Pro or Ultra subscriber in the Gemini app and AI Mode in Search. There is no free tier on the API.

How large is Gemini 4 Argon's context window?

It has a unified context window of 1,048,576 tokens, shared by input, thinking and output. Default maximum output is 65,536 tokens, rising to around 1 million through long decode.

Is Gemini 4 Argon worth building on before 30 November?

It is worth prototyping before 30 November if your workload involves long documents, large codebases or complex reasoning. Use real data, measure per-task cost at both prices, and commit only if your own evaluation supports it.

What is the cheapest way to run long-context work on Argon?

Reuse stable reference material for 95% off cached input, use the lowest thinking level that holds accuracy, cap output length, and budget per task. Output tokens cost more than input tokens, so concise outputs save most.