Gemini Models: Argon, Flash and Flash-Lite Compared

Google's current Gemini API lineup has one flagship, Gemini 4 Argon ($2 input and $10 output per million tokens until 30 November 2026, then $4 and $20), and a Flash tier below it: 3.8 Flash at $0.75 and $3.75, and 3.5 Flash-Lite at $0.30 and $2.50. They all take text, image, video, audio and PDF input and a context window of about 1M tokens. Checked on 31 October 2026.

The short version

Prices below are Google's list prices for the Standard tier, in US dollars per million tokens, taken from the Gemini API pricing page today. Two of them change soon, and we flag that every time it matters.

The current lineup

Google's models page lists the Gemini 4 family and the Gemini 3 family separately. In practice you choose between three text models and a few older or specialised ones.

Two details from the model pages are easy to miss. Argon supports implicit caching only, so you cannot create explicit caches for it. And 3.8 Flash does not accept a "minimal" thinking level: it supports low, medium and high, and a request for minimal returns an error.

Older and specialised models still on the page

We cover the Pro line in more detail on the Gemini Pro page and the Flash line on the Gemini Flash page.

Maker, licence and open weights

Gemini is made by Google. The Gemini API models are closed-weight: there is no download, no self-hosting and no weights licence to read. You get access under Google's API terms, either directly through the Gemini API and Google AI Studio or through Google Cloud. Google's open-weights family is a separate product called Gemma, so if the reason you are reading this is a need to run a model on your own hardware, jump to the Gemma page or our open source LLM overview.

The practical consequence is simple. You cannot pin a Gemini version forever. Google publishes model version patterns, and the stable IDs above point at a stable release, but older models are eventually retired. Build your agent so that the model ID is a setting, not a constant.

Context window and output

Every text model in the table has a 1,048,576-token window. That is roughly 750,000 English words, using the rule of thumb on most vendor pages that 1,000 tokens is about 750 words. Argon's window is a unified limit shared between input and output, which matters because of long decode.

Long decode is the new feature worth understanding. By default Argon returns up to 65,536 tokens in a synchronous call. With long decode enabled, Google lists the maximum output as up to 1,048,576 tokens. That makes it possible to ask for a very large generated artefact, such as a full code migration, in one response. It also means one request can be expensive. At $10 per million output tokens (the price until 30 November 2026), a full 1M-token generation costs $10 in output alone. After 1 December it would cost $20.

Input pricing on the Pro preview model steps up at 200K tokens, while the Flash and Argon pages list a single rate. If you plan to fill a long window, read the row for your model rather than assuming the Pro pattern.

Prices in practice

List prices are easier to compare when you push the same job through each model. Take a job that reads 10 million input tokens and writes 2 million output tokens, such as a repository-wide review run over a few hundred files. The sums below use only the Standard tier list prices above, with no caching and no batch discount.

Thinking tokens are billed as output. Google's pricing page says output price includes thinking tokens, so a model set to high thinking can spend a lot more than the visible answer suggests. If your run costs more than the table predicts, that is the first place to look.

Ways to pay less

Google's pricing page lists several tiers for the same models, and they matter more than shaving a model tier.

Batch and Flex are half the Standard price on all three models. Priority costs 1.8 times Standard on Argon and 3.8 Flash, and 1.8 times on Flash-Lite as well, going by the listed numbers. Context caching is cheaper again: $0.10 per million cached tokens on Argon until 30 November (then $0.20), $0.075 on 3.8 Flash until 31 December (then $0.15), and $0.03 on 3.5 Flash-Lite. Flash models also charge a storage price per million tokens per hour for explicit caches, and Argon has no explicit caching at all.

Search grounding is billed separately: 5,000 free search requests a month, shared across Gemini 3 and newer models, then $14 per 1,000 requests, with a daily cap of 1,500.

Where it is served

Gemini models are served by Google: through the Gemini API and AI Studio for developers, and through Google Cloud for enterprises, which Google's docs name the Gemini Enterprise Agent Platform. They are not on our list of open-weights hosts. None of the pages under /hosting/ on this site, such as Runpod or Modal, run Gemini, because those sell GPU time for models you can download.

If you want one API key that reaches Gemini and other vendors, a gateway is the usual answer. We compare those in OpenRouter alternatives. For a straight look at what each vendor charges for an input and output token, see LLM API prices and Gemini API pricing.

Which coding agents can use it

Google's own agents are the obvious home: Gemini CLI and Antigravity. Google's Gemini API docs also carry a "Coding agent setup" page and an Antigravity agent page.

Beyond Google's tools, any agent that lets you bring an API key and choose a model can call the Gemini API. Cursor, Cline and OpenCode are examples of agents with model pickers, and Gemini's docs list an OpenAI-compatibility endpoint, which many agents accept as a custom base URL. Claude Code is built around Anthropic's own models, so routing it to Gemini means a gateway and, in our reading, less certain behaviour. We say more on that in our DeepSeek versus Claude page, which covers the same problem for another non-Anthropic model.

For the choice of tool itself, choose an AI coding agent is the right starting point.

Benchmarks

Google's models pages for these three models point to model cards rather than printing scores, and we did not copy any figures from third-party leaderboards. Treat any "Gemini beats X" claim you see as unverified. When you compare models, run your own repository and your own tasks. Vendor-reported tables are marketing by construction, and ours would be too if we copied them without the source and a label.

Privacy and data use

The pricing page has a row called "Used to improve our products". For Argon, 3.8 Flash on the paid tier and 3.5 Flash-Lite on the paid tier it says No. On the free tier of 3.8 Flash and 3.6 Flash it says Yes. The rule is straightforward: if the code or data is private, use the paid tier. The free tier is for experiments and public code.

Argon has no free tier, so the question does not arise there. Data location, retention periods and zero-retention terms are separate and are not on the pricing page.

Limits and gotchas

• Price changes are scheduled. Argon doubles on 1 December 2026 and 3.8 Flash doubles on 1 January 2027. Budget for the later number.
• Argon has no free tier, and no explicit caching.
• Minimal thinking returns an error on 3.8 Flash. Use low, medium or high.
• Pro is still a preview. The 3.1 Pro model is labelled preview, and Google may change preview models with less notice than stable ones.
• Older 2.5 models are gated. New projects are steered to 3.5 Flash-Lite or 3.8 Flash.
• Long decode can be costly. Up to 1M output tokens in one call is a feature and a bill.
• Grounding adds cost. Past the free monthly allowance, each search is billed even if the answer was cached elsewhere.

Our take

Treat Gemini as a three-step ladder. Start your agent on 3.8 Flash. It is the model Google itself points to for long-horizon software engineering, and at $0.75 and $3.75 it costs about a third of what Argon will cost after 1 December. Move specific hard steps, such as a planning pass or a stubborn bug, up to Argon. Move bulk, low-risk steps, such as summarising files or extracting fields, down to Flash-Lite.

The cheap-until-date prices deserve a calendar entry. If you pick Argon today because $2 and $10 looks good, your bill doubles on 1 December with no code change. That is not a reason to avoid it. It is a reason to measure how much better it is on your own tasks before the discount ends.

Gemini's real selling point in this lineup is breadth: text, image, video, audio and PDF in on every model, a 1M window on all of them, and voice, image and video models under the same API key. If you only write code, that breadth matters less, and the price per token and the agent you use decide the outcome.

Who should pick what

Related pages

• Gemini Flash and Gemini Pro for each line in depth
• Gemini API pricing for every tier and tool charge
• Claude models and GPT models for the other two big closed families
• Compare AI models side by side for the interactive table