Gemini Pro and Argon: Specs, Prices and Limits

Google's top Gemini models are Gemini 4 Argon, a stable flagship at $2 input and $10 output per million tokens until 30 November 2026 and $4 and $20 after, and Gemini 3.1 Pro Preview, a preview at $2 and $12 for prompts up to 200K tokens. Both take text, images, video, audio and PDFs, and both have a 1M-token window. Neither has a free API tier. Checked on 31 October 2026.

First, which one is "Pro"?

The name causes confusion, so here is exactly what Google's models page says. Gemini 3.1 Pro (gemini-3.1-pro-preview) is the Pro model. It is labelled Preview. Gemini 4 Argon (gemini-4-argon) is described as Google's flagship reasoning model, is labelled Stable, and its page was updated in October 2026. The word "Pro" does not appear in Argon's name or description on the pages we read, and we found no Google statement that Argon succeeds the Pro line.

So when someone writes "Gemini Pro" in a tutorial from earlier this year, they mean the Pro line of Gemini 2.5 or 3. When you build today, the realistic choice is between Argon and 3.1 Pro Preview, and the specs and prices below are what separate them. For the cheaper tier, see the Gemini Flash page, and for the full family see Gemini models.

Specs side by side

Argon's capability list includes agentic video understanding (defaulting to static), code execution, computer use, file search, function calling, Maps and Search grounding, structured outputs and URL context. 3.1 Pro Preview lists the same set minus computer use, and its file search is marked "AI Studio only." That second detail matters if you plan to build file search into an agent through the API.

Prices

All amounts are US dollars per million tokens on the paid tier, taken from the Gemini API pricing page today.

Argon's batch, flex and priority prices also double on 1 December, following the Standard rate. Search grounding is the same for both: 5,000 free requests a month across Gemini 3 and newer models, then $14 per 1,000, capped at 1,500 a day. Explicit-cache storage on 3.1 Pro is $4.50 per million tokens per hour. Argon supports implicit caching only, so there is no storage line for it.

Output prices include thinking tokens on both models, so a response that thinks for a long time bills for all of it.

Worked example: a 500K-token repository prompt

Suppose an agent sends a whole repository, 500,000 tokens, and gets 20,000 tokens back. On 3.1 Pro the prompt is over 200K, so the higher rate applies to every token in the request The pricing page lists the higher rate as "prompts > 200k tokens" and does not spell out whether it covers the whole prompt, so budget as if it does.

The comparison flips after 1 December. Until then Argon is cheaper than Pro for a long prompt, and after it Argon is about the same price as 3.1 Pro on this job. If you use the repository as the opening of every request, caching changes the picture more than the model choice does: cached input on Argon is $0.10 per million through November, against $2.00 for fresh input.

Maker, licence and open weights

Google makes both models. They are closed-weight and available only through Google's APIs and Google Cloud. There is no model licence to read and no download. If you need Google's open weights, use Gemma and read the Gemma page. For the full list of models you can host yourself, see our open source LLM overview.

Where it is served

Directly from Google, through the Gemini API, AI Studio and Google Cloud. The hosts under /hosting/ on this site sell GPU time for downloadable models, so Runpod, Lambda and Modal do not apply to Argon or Pro. We did not check whether other clouds' catalogues carry Gemini models. A multi-vendor gateway is the other route, and we weigh those in OpenRouter alternatives.

Coding agents that can use it

Gemini CLI and Antigravity are Google's own agents. Google's docs mention an Antigravity agent and a coding-agent setup page in the Gemini API section. Antigravity's model docs list Gemini 4 Argon on Google AI Pro and Ultra and not on the free plan, which has Gemini 3.8 Flash. We did not check Gemini CLI's model list. Any agent that accepts a Gemini API key or an OpenAI-compatible endpoint can reach these models, subject to its own model list. For a wider comparison of agents, read choose an AI coding agent.

Benchmarks

Google's model pages for Argon and 3.1 Pro Preview describe use cases and link to model cards. They do not print scores in the pages we fetched, and we did not copy numbers from leaderboards. A vendor card reports the vendor's own setup, so when we add scores we will label them vendor-reported and name the source.

Calling Argon

The model ID is the only thing that changes in a call. This is the shape of a request with the Google GenAI SDK for Python, as a sketch rather than a tested recipe. Check the current quickstart in Google's docs before you copy it.

from google import genai

client = genai.Client()  # reads the API key from the environment
response = client.models.generate_content(
    model="gemini-4-argon",
    contents="Find the bug in this function and explain it.",
)
print(response.text)

Two habits save money. First, keep the model ID in configuration so a price change or a retirement is a settings change. Second, log the token counts from every response, including thinking tokens, so you can compare real spend against the sums on this page.

Moving from 3.1 Pro Preview to Argon

If you already run on the Pro preview, here is a short checklist for the move.

1. Price. Compare on your own traffic. Argon is cheaper on long prompts until 30 November and similar after.
2. Thinking levels. Argon lists low, medium and high. Set one explicitly rather than relying on a default.
3. Caching. Argon supports implicit caching only. If you built explicit caches for 3.1 Pro, they do not carry over, and the storage charge goes away with them.
4. File search. If you used file search from AI Studio only, check how Argon exposes it, since its page lists file search as supported without that limit.
5. Tools. Argon lists computer use. If your agent drives a browser, you may now be able to do it on the same model.
6. Output ceilings. Both default to 65,536 tokens. Turn on long decode only for jobs that need more.
7. Tests. Run your regression set on both models and compare passes, retries and spend before you cut over.

Picking a tier

Priority on Argon costs $3.60 and $18 until 30 November, and double after. Google's Priority page says it is designed for second-level response times, has non-sheddable traffic, starts at 0.3x the standard rate limit, and moves overflow requests to Standard processing at Standard rates instead of failing them.

Long decode, in plain terms

Argon's headline feature is long decode, which Google describes as output up to 1M tokens. By default a synchronous request returns at most 65,536 tokens. With long decode on, the limit rises to 1,048,576. That is the only place a Gemini model writes more than 65,536 tokens in one response among the models we checked.

Use it when the job is one huge artefact: a converted codebase, a long document set, a large test suite. Do not use it when the job is many small steps. An agent that edits a file at a time never needs it, and a runaway generation can cost real money. At the pre-December Argon output price of $10 per million, a full 1M-token reply costs $10. After 1 December it is $20. Set a hard cap in your client, and prefer streaming so you can stop a bad run early. The long decode page says it runs on the Interactions API with background: true, store: true and max_output_tokens above 65,536, covers only Gemini 4 Argon, and does not mention Batch.

Privacy

The pricing page's "Used to improve our products" row says No for Argon and for 3.1 Pro Preview on the paid tier, and neither has a free tier. That is the cleanest posture in the Gemini lineup for private code. Retention periods, processing regions and zero-retention options are separate topics. Google's API terms say it logs prompts and responses for a limited period solely to detect and prevent violations, and that data may be stored transiently or cached in any country where Google or its agents have facilities.

Limits and gotchas

• Argon doubles on 1 December 2026. Plan for $4 and $20.
• 3.1 Pro is a preview. Google can change or retire previews faster than stable models.
• Long prompts cost more on 3.1 Pro. The rate steps up past 200K tokens. Argon's page shows one rate.
• Argon has no explicit caching. Rely on implicit caching and stable prompt prefixes.
• File search on 3.1 Pro is AI Studio only, per its model page.
• No computer use on 3.1 Pro Preview in its capability list, while Argon lists it as supported.
• Output limits differ. 65,536 tokens by default on both. Only Argon lists long decode beyond that.

Our take

If you are starting a new agent today and want Google's best, test Argon, and test it before 1 December so you know what you are buying at the doubled price. The model page language, flagship status and the long-decode feature all point to it as the model Google wants developers on. 3.1 Pro Preview is the one to leave alone for new work unless you need something specific from it: it is a preview, it has the 200K price step, and its capability list is a little narrower.

At the doubled price, Argon's list price of $4 and $20 matches Claude Opus 5.5 exactly, so price will not settle the choice. What settles it is whether the model does your tasks well. Compare it against Claude Opus and the GPT models on your own repository, and count passes per dollar, not tokens per dollar.

Who should pick what

Related pages

• Gemini models for the full lineup
• Gemini Flash for the cheaper tier
• Gemini API pricing for every tier and tool charge
• Claude Opus for the closest premium rival on list price