Gemini 4 was one of the top AI search terms on X in our sample, with about 14,400 posts, though interest is falling since launch week. Most builders have one question: is Argon worth switching to, and how long will the cheap price last?
What Google shipped
Gemini 4 Argon is Google's new frontier model, released on Oct 13, 2026. Google pitches it at "deep work": long, multi-step tasks in coding, finance, legal, cybersecurity and science.
Where you can use it:
| Who | Where |
|---|---|
| Developers | Gemini API paid tier in Google AI Studio and Android Studio |
| AI Ultra subscribers | Google Antigravity and AI Studio |
| AI Pro and Ultra subscribers | Gemini app and AI Mode in Search |
| Enterprises | Gemini for business, the API, Antigravity |
The pricing, and the Nov 30 cliff
| Until Nov 30, 2026 | After Nov 30, 2026 | |
|---|---|---|
| Input | $2 / M tokens | $4 / M tokens |
| Output | $10 / M tokens | $20 / M tokens |
| Cached input | 95% off | Not stated in the launch post |
The launch price is officially "introductory". Google's footnote says that after it expires on Nov 30, 2026, $4 input and $20 output apply.
If you're building cost models, use the post-November numbers. A prototype that looks cheap in October will cost twice as much per token in December.
The caching discount is the big lever. At 95% off, a large repeated prefix costs almost nothing to re-send. Think of a long system prompt, a codebase map or a document set. Our prompt caching guide explains how to structure prompts so the cache actually hits.
The 1M unified token limit
Google describes "an industry-leading 1M unified input-output token limit via the Gemini API". Input and output share one budget. Google also mentions a new continuation protocol for long jobs and points developers to its docs for setup.
In practice, you don't have to choose upfront between a huge input and a long output. You still pay for every token, though, so a full 1M-token run isn't cheap even at intro prices. Our context windows explainer covers why "fits in the window" doesn't mean "the model will use it well."
What Google says it's good at
Treat all of these as Google's claims:
- #1 on the Vals Index, which scores models on finance, coding, legal and tax work.
- In Antigravity, Google says Argon raised agentic code submissions by more than 35% and cut rollbacks by nearly a third.
- Real refactors. Google says Argon helped the Chrome team split a roughly 6-million-line legacy folder out of Chromium.
- Performance work. Google reports 40% faster in-memory session prep in its ADK Python framework, and a 2.85x vLLM inference speedup for Gemma 4 E2B on an A100.
These are useful signals about where Google aimed the model: long refactors and performance work, not chat.
When to pick Argon
Good fit: - Long agent runs with a big, stable context you can cache. - Large refactors and migrations, where Google focused its examples. - Teams already on Antigravity or Android Studio.
Think twice: - High-volume, simple calls. A cheaper small model will usually win on cost. See our LLM API pricing comparison. - Budgets that can't absorb the price doubling in December.
How it compares on price
For context, OpenAI's GPT-6 Astra lists at $10 input and $50 output per million tokens. Even at its post-November price, Argon is cheaper per token. Per-token price isn't per-task cost, though. A model that finishes in fewer steps can come out cheaper overall. Run the same five real tasks on each candidate and compare the bills.
To keep switching easy, keep your prompts and tooling model-agnostic. Our guide to picking a coding agent shows how.
