Models
Gemini 4 Argon and Flash: What Google Offers for the Price
Gemini is Google DeepMind’s model family. Its main advantages are huge context windows, multimodal input for text, images, audio, and video in one request, and aggressive pricing. In 2026, Gemini 4 Argon is the flagship, while Flash is the workhorse with a context of up to 10 million tokens. Here is what that means in practice.
Current Gemini models
Argon launched on September 30, 2026, with two unusual features. First is its output window: up to one million tokens in a single generation, compared with roughly 64,000 in the previous generation. Google is betting on long trajectories: an agent task that once needed dozens of calls may fit into one. Second is access: the model was initially rolled out in stages to vetted cybersecurity specialists in the Fairwind program, developers, and AI Ultra subscribers after safety checks.
Flash is built for scale: 10 million context tokens at $0.15 per million. That can fit a 500,000-line codebase or an hour of video into one request.
| Model | Context | Price per 1M tokens |
|---|---|---|
| Gemini 4 Argon (flagship) | 1M input / up to 1M output | $2 input / $10 output at launch, then $4 / $20 |
| Gemini 4.0 Flash | Up to 10M tokens | $0.15 input / $0.60 output |
| Gemini 3.x (previous generation) | Up to 2M tokens | Lower cost, but less capable |
How capable is it?
Benchmarks put Argon alongside top flagship models, with strong results in real-world software development (DeepSWE), agent automation, and long-video understanding. One claim is that it can work with an 800,000-line codebase as one project rather than as separate fragments.
Multimodality has long been a Gemini advantage. The model accepts images, audio, and videos several hours long without converting them to text first. This works out of the box for meeting recordings, lectures, and interface screenshots, while many competitors either cannot accept video or limit it to a few minutes.
Safety and refusals
With Argon, Google took a “safety first, access second” approach: the public rollout happened in waves, so availability may differ from what was announced by the time you read this. Topics such as cyberattacks and harmful content are filtered at several levels: the model, service, and regional rules. Behavior can therefore differ even between accounts.
Refusals on everyday requests are uncommon. A Gemini-specific detail is that API moderation is strict by default: safety filters are configured separately, and high blocking levels can stop harmless answers. If responses are cut off, check the safety settings.
What does a task cost?
Gemini prices per token have traditionally been among the lowest. For example, 10 million input tokens with Flash cost about $1.50; a comparable context can cost tens of dollars with competitors. At launch, cached Argon input was another 5–95% cheaper, depending on the use case.
Keep in mind that cheap input does not mean a cheap task. Long answers are billed as output, and Flash output costs four times as much as input. Agent loops that read, reason, and answer again can spend heavily on repeated reads, so caching matters.
How Gemini works
- Flash supports up to 10 million context tokens, but test how well it retains details near the end of the window on your task.
- Native, MCP-compatible tool calls are built into the model, reducing the amount of agent scaffolding needed.
- Built-in Google tools include Search, Maps, code execution, and URL context.
- Flash supports multimodal input, including up to three hours of video in one request.
- Limit: flagship Argon access is staged, and not every mode is immediately available in the consumer app.
Who should use Gemini?
Gemini is a strong choice for large inputs: video analysis, huge documents, or entire repositories. It is also a sensible low-cost default when flagship-level quality is more than your routine needs.
For clean code and polished writing, many people prefer Claude or GPT. Gemini is stronger when multimodality and scale matter. The most reliable way to choose is to run the same prompt through each model; in Any LLM chat, you can switch models before each question.
Frequently asked questions
Is Gemini free?
The Gemini app has a free tier, and AI Studio has usage limits. Full access to large contexts and flagship models is paid.
How much is 10 million tokens?
Roughly 7–8 million words, or hundreds of average-length books. In practice, it could be a large codebase or several hours of video.
Why does Gemini refuse more often than other models?
API safety settings or regional rules may be responsible. Check the safety settings and the wording of your request.
Is Argon available to everyone?
No. The rollout is staged: cybersecurity specialists and Ultra subscribers first, then other users. Check Google’s pages for current availability.
Should I choose Gemini Flash or GPT Luna?
Both suit high-volume routine work. Flash offers a larger context and multimodality; Luna costs less for simple text tasks.